neuro-simulator 0.1.3__py3-none-any.whl → 0.2.1__py3-none-any.whl

neuro_simulator/agent/core.py

@@ -1,18 +1,25 @@
-# agent/core.py
+# neuro_simulator/agent/core.py
 """
-Core module for the Neuro Simulator Agent
+Core module for the Neuro Simulator's built-in agent.
 """
 
-import os
-import json
 import asyncio
-from typing import Dict, List, Any, Optional
-from datetime import datetime
-import sys
+import json
 import logging
+import re
+import sys
+from pathlib import Path
+from datetime import datetime
+from typing import Any, Dict, List, Optional
 
-# Import the shared log queue from the main log_handler
-from ..log_handler import agent_log_queue, QueueLogHandler
+# Updated imports for the new structure
+from ..utils.logging import QueueLogHandler, agent_log_queue
+from ..utils.websocket import connection_manager
+
+# --- Agent-specific imports ---
+from .llm import LLMClient
+from .memory.manager import MemoryManager
+from .tools.core import ToolManager
 
 # Create a logger for the agent
 agent_logger = logging.getLogger("neuro_agent")
@@ -20,475 +27,175 @@ agent_logger.setLevel(logging.DEBUG)
 
 # Configure agent logging to use the shared queue
 def configure_agent_logging():
-    """Configure agent logging to use the shared agent_log_queue"""
-    # Create a handler for the agent queue
-    agent_queue_handler = QueueLogHandler(agent_log_queue)
-    formatter = logging.Formatter('%(asctime)s - [AGENT] - %(levelname)s - %(message)s', datefmt='%H:%M:%S')
-    agent_queue_handler.setFormatter(formatter)
-    
-    # Clear any existing handlers
+    """Configure agent logging to use the shared agent_log_queue."""
     if agent_logger.hasHandlers():
         agent_logger.handlers.clear()
     
-    # Add the queue handler
+    agent_queue_handler = QueueLogHandler(agent_log_queue)
+    # Use the same format as the server for consistency
+    formatter = logging.Formatter('%(asctime)s - [%(name)-24s] - %(levelname)-8s - %(message)s', datefmt='%H:%M:%S')
+    agent_queue_handler.setFormatter(formatter)
     agent_logger.addHandler(agent_queue_handler)
-    agent_logger.propagate = False  # Prevent logs from propagating to root logger
-    
-    print("Agent日志系统已配置，将日志输出到 agent_log_queue。")
+    agent_logger.propagate = False
+    agent_logger.info("Agent logging configured to use agent_log_queue.")
 
-# Configure agent logging when module is imported
 configure_agent_logging()
 
 class Agent:
-    """Main Agent class that integrates LLM, memory, and tools"""
+    """Main Agent class that integrates LLM, memory, and tools. This is the concrete implementation."""
     
     def __init__(self, working_dir: str = None):
-        # Lazy imports to avoid circular dependencies
-        from .memory.manager import MemoryManager
-        from .tools.core import ToolManager
-        from .llm import LLMClient
-        
         self.memory_manager = MemoryManager(working_dir)
         self.tool_manager = ToolManager(self.memory_manager)
         self.llm_client = LLMClient()
         self._initialized = False
-        
-        # Log agent initialization
-        agent_logger.info("Agent initialized")
+        agent_logger.info("Agent instance created.")
         agent_logger.debug(f"Agent working directory: {working_dir}")
         
     async def initialize(self):
-        """Initialize the agent, loading any persistent memory"""
+        """Initialize the agent, loading any persistent memory."""
         if not self._initialized:
-            agent_logger.info("Initializing agent memory manager")
+            agent_logger.info("Initializing agent memory manager...")
             await self.memory_manager.initialize()
             self._initialized = True
-            agent_logger.info("Agent initialized successfully")
+            agent_logger.info("Agent initialized successfully.")
         
     async def reset_all_memory(self):
-        """Reset all agent memory types"""
-        # Reset temp memory
+        """Reset all agent memory types."""
         await self.memory_manager.reset_temp_memory()
-        
-        # Reset context (dialog history)
         await self.memory_manager.reset_context()
-        
-        agent_logger.info("All agent memory reset successfully")
-        print("All agent memory reset successfully")
-        
-    async def reset_memory(self):
-        """Reset agent temp memory (alias for backward compatibility)"""
-        await self.reset_all_memory()
+        agent_logger.info("All agent memory has been reset.")
         
     async def process_messages(self, messages: List[Dict[str, str]]) -> Dict[str, Any]:
-        """
-        Process incoming messages and generate a response
-        
-        Args:
-            messages: List of message dictionaries with 'username' and 'text' keys
-            
-        Returns:
-            Dictionary containing processing details including tool executions and final response
-        """
-        # Ensure agent is initialized
+        """Process incoming messages and generate a response with tool usage."""
         await self.initialize()
-        
-        agent_logger.info(f"Processing {len(messages)} messages")
-        
-        # Add messages to context
+        agent_logger.info(f"Processing {len(messages)} messages.")
+
         for msg in messages:
             content = f"{msg['username']}: {msg['text']}"
             await self.memory_manager.add_context_entry("user", content)
-            agent_logger.debug(f"Added message to context: {content}")
-            
-        # Send context update via WebSocket after adding user messages
-        from ..websocket_manager import connection_manager
+        
         context_messages = await self.memory_manager.get_recent_context()
-        await connection_manager.broadcast({
-            "type": "agent_context",
-            "action": "update",
-            "messages": context_messages
-        })
+        await connection_manager.broadcast({"type": "agent_context", "action": "update", "messages": context_messages})
         
-        # Add detailed context entry for the start of processing
         processing_entry_id = await self.memory_manager.add_detailed_context_entry(
-            input_messages=messages,
-            prompt="Processing started",
-            llm_response="",
-            tool_executions=[],
-            final_response="Processing started"
+            input_messages=messages, prompt="Processing started", llm_response="",
+            tool_executions=[], final_response="Processing started"
         )
             
-        # Get full context for LLM
         context = await self.memory_manager.get_full_context()
         tool_descriptions = self.tool_manager.get_tool_descriptions()
         
-        # Get last agent response to avoid repetition
-        last_response = await self.memory_manager.get_last_agent_response()
-        
-        # Create LLM prompt with context and tools
-        prompt = f"""You are {self.memory_manager.init_memory.get('name', 'Neuro-Sama')}, an AI VTuber.
-Your personality: {self.memory_manager.init_memory.get('personality', 'Friendly and curious')}
-
-=== CONTEXT ===
-{context}
-
-=== AVAILABLE TOOLS ===
-{tool_descriptions}
-
-=== INSTRUCTIONS ===
-Process the user messages and respond appropriately. You can use tools to manage memory or output responses.
-When you want to speak to the user, use the 'speak' tool with your response as the text parameter.
-When you want to update memory, use the appropriate memory management tools.
-You are fully responsible for managing your own memory. Use the memory tools proactively when you need to:
-- Remember important information from the conversation
-- Update your knowledge or personality
-- Store observations about users or events
-- Retrieve relevant information to inform your responses
-Always think about whether you need to use tools before responding.
+        # --- CORRECTED HISTORY GATHERING ---
+        recent_history = await self.memory_manager.get_detailed_context_history()
+        assistant_responses = []
+        for entry in reversed(recent_history):
+            if entry.get("type") == "llm_interaction":
+                for tool in entry.get("tool_executions", []):
+                    if tool.get("name") == "speak" and tool.get("result"):
+                        assistant_responses.append(tool["result"])
 
-IMPORTANT GUIDELINES:
-- Be creative and engaging in your responses
-- Keep responses concise and conversational
-- Maintain your character's personality
-- Pay close attention to the conversation history in the context to understand the flow of the dialogue
-- Respond to the most recent user messages while considering the overall context of the conversation
+        # Create LLM prompt from template
+        template_path = Path(self.memory_manager.memory_dir).parent / "prompt_template.txt"
+        with open(template_path, 'r', encoding='utf-8') as f:
+            prompt_template = f.read()
 
-=== YOUR SPEAK HISTORY ===
-"""
+        recent_speak_history_text = "\n".join([f"- {response}" for response in assistant_responses[:5]]) if assistant_responses else "You haven't said anything yet."
+        user_messages_text = "\n".join([f"{msg['username']}: {msg['text']}" for msg in messages])
 
-        # Add assistant's recent responses to the prompt
-        # Get the recent context and extract speak results from llm_interaction entries
-        recent_context = await self.memory_manager.get_recent_context(100)  # Get more entries to filter
-        assistant_responses = []
-        
-        # Filter for llm_interaction entries with role: "assistant" and extract speak results
-        for entry in reversed(recent_context):  # Reverse to get newest first
-            if entry.get("type") == "llm_interaction" and entry.get("role") == "assistant":
-                tool_executions = entry.get("tool_executions", [])
-                for tool_execution in tool_executions:
-                    if tool_execution.get("name") == "speak":
-                        result = tool_execution.get("result")
-                        if result:
-                            assistant_responses.append(result)
-                    
-        # Add up to 64 most recent assistant responses
-        for i, response in enumerate(assistant_responses[:64]):
-            prompt += f"{i+1}. {response}\n"
-            
-        prompt += f"\nUser messages to respond to:\n"
-        
-        for msg in messages:
-            prompt += f"{msg['username']}: {msg['text']}\n"
-            
-        prompt += "\nYour response (use tools as needed):"
-        
-        agent_logger.debug("Sending prompt to LLM")
+        prompt = prompt_template.format(
+            full_context=context,
+            tool_descriptions=tool_descriptions,
+            recent_speak_history=recent_speak_history_text,
+            user_messages=user_messages_text
+        )
         
-        # Add detailed context entry for the prompt
         await self.memory_manager.add_detailed_context_entry(
-            input_messages=messages,
-            prompt=prompt,
-            llm_response="",
-            tool_executions=[],
-            final_response="Prompt sent to LLM",
-            entry_id=processing_entry_id
+            input_messages=messages, prompt=prompt, llm_response="", tool_executions=[],
+            final_response="Prompt sent to LLM", entry_id=processing_entry_id
         )
         
-        # Generate response using LLM
-        response = await self.llm_client.generate(prompt)
-        agent_logger.debug(f"LLM response received: {response[:100] if response else 'None'}...")
+        response_text = await self.llm_client.generate(prompt)
+        agent_logger.debug(f"LLM raw response: {response_text[:100] if response_text else 'None'}...")
         
-        # Add detailed context entry for the LLM response
         await self.memory_manager.add_detailed_context_entry(
-            input_messages=messages,
-            prompt=prompt,
-            llm_response=response,
-            tool_executions=[],
-            final_response="LLM response received",
-            entry_id=processing_entry_id
+            input_messages=messages, prompt=prompt, llm_response=response_text, tool_executions=[],
+            final_response="LLM response received", entry_id=processing_entry_id
         )
         
-        # Parse the response to handle tool calls
-        # This is a simplified parser - in a full implementation, you would use a more robust method
         processing_result = {
-            "input_messages": messages,
-            "llm_response": response,
-            "tool_executions": [],
-            "final_response": ""
+            "input_messages": messages, "llm_response": response_text,
+            "tool_executions": [], "final_response": ""
         }
         
-        # Extract tool calls from the response
-        # Look for tool calls in the response
-        lines = response.split('\n') if response else []
-        i = 0
-        json_buffer = ""  # Buffer to accumulate multi-line JSON
-        in_json_block = False  # Flag to track if we're inside a JSON block
-        
-        while i < len(lines):
-            line = lines[i].strip()
-            agent_logger.debug(f"Parsing line: {line}")
-            
-            # Handle JSON blocks
-            if line.startswith('```json'):
-                in_json_block = True
-                json_buffer = line + '\n'
-            elif line == '```' and in_json_block:
-                # End of JSON block
-                json_buffer += line
-                in_json_block = False
-                # Process the complete JSON block
-                tool_call = self._parse_tool_call(json_buffer)
-                if tool_call:
-                    agent_logger.info(f"Executing tool: {tool_call['name']}")
-                    await self._execute_parsed_tool(tool_call, processing_result)
-                    # Update detailed context entry for tool execution
-                    await self.memory_manager.add_detailed_context_entry(
-                        input_messages=messages,
-                        prompt=prompt,
-                        llm_response=response,
-                        tool_executions=processing_result["tool_executions"].copy(),  # Pass a copy of current tool executions
-                        final_response=f"Executed tool: {tool_call['name']}",
-                        entry_id=processing_entry_id
-                    )
-                else:
-                    agent_logger.warning(f"Failed to parse tool call from JSON block: {json_buffer}")
-            elif in_json_block:
-                # Accumulate lines for JSON block
-                json_buffer += line + '\n'
-            else:
-                # Check if line contains a tool call
-                if any(line.startswith(prefix) for prefix in ["get_", "create_", "update_", "delete_", "add_", "remove_", "speak("]):
-                    # Parse tool call
-                    tool_call = self._parse_tool_call(line)
-                    if tool_call:
-                        agent_logger.info(f"Executing tool: {tool_call['name']}")
-                        await self._execute_parsed_tool(tool_call, processing_result)
-                        # Update detailed context entry for tool execution
-                        await self.memory_manager.add_detailed_context_entry(
-                            input_messages=messages,
-                            prompt=prompt,
-                            llm_response=response,
-                            tool_executions=processing_result["tool_executions"].copy(),  # Pass a copy of current tool executions
-                            final_response=f"Executed tool: {tool_call['name']}",
-                            entry_id=processing_entry_id
-                        )
-                    else:
-                        agent_logger.warning(f"Failed to parse tool call from line: {line}")
-            i += 1
-            
-        # If we're still in a JSON block at the end, process it
-        if in_json_block and json_buffer:
-            tool_call = self._parse_tool_call(json_buffer)
-            if tool_call:
+        if response_text:
+            tool_calls = self._parse_tool_calls(response_text)
+            for tool_call in tool_calls:
                 agent_logger.info(f"Executing tool: {tool_call['name']}")
                 await self._execute_parsed_tool(tool_call, processing_result)
-                # Update detailed context entry for tool execution
-                await self.memory_manager.add_detailed_context_entry(
-                    input_messages=messages,
-                    prompt=prompt,
-                    llm_response=response,
-                    tool_executions=processing_result["tool_executions"].copy(),  # Pass a copy of current tool executions
-                    final_response=f"Executed tool: {tool_call['name']}",
-                    entry_id=processing_entry_id
-                )
-            else:
-                agent_logger.warning(f"Failed to parse tool call from incomplete JSON block: {json_buffer}")
-        
-        # If we have a final response, we don't need to add it to context here
-        # because add_detailed_context_entry will add a short entry for us
-        # if processing_result["final_response"]:
-        #     await self.memory_manager.add_context_entry("assistant", processing_result["final_response"])
-            
-        # Update the detailed context entry with final LLM interaction details
+
         await self.memory_manager.add_detailed_context_entry(
-            input_messages=messages,
-            prompt=prompt,
-            llm_response=response,
+            input_messages=messages, prompt=prompt, llm_response=response_text,
             tool_executions=processing_result["tool_executions"],
-            final_response=processing_result["final_response"],
-            entry_id=processing_entry_id
+            final_response=processing_result["final_response"], entry_id=processing_entry_id
         )
             
-        # Send context update via WebSocket
-        from ..websocket_manager import connection_manager
-        context_messages = await self.memory_manager.get_recent_context()
-        await connection_manager.broadcast({
-            "type": "agent_context",
-            "action": "update",
-            "messages": context_messages
-        })
+        final_context = await self.memory_manager.get_recent_context()
+        await connection_manager.broadcast({"type": "agent_context", "action": "update", "messages": final_context})
             
-        agent_logger.info("Message processing completed")
+        agent_logger.info("Message processing completed.")
         return processing_result
         
     async def _execute_parsed_tool(self, tool_call: Dict[str, Any], processing_result: Dict[str, Any]):
-        """Execute a parsed tool call and update processing result"""
-        # Only prevent duplicate speak tool executions to avoid repeated responses
-        if tool_call["name"] == "speak":
-            for executed_tool in processing_result["tool_executions"]:
-                if (executed_tool["name"] == "speak" and 
-                    executed_tool["params"].get("text") == tool_call["params"].get("text")):
-                    agent_logger.debug(f"Skipping duplicate speak tool execution: {tool_call['params'].get('text')}")
-                    return
-        
-        # Execute the tool
+        """Execute a parsed tool call and update processing result."""
         try:
             tool_result = await self.execute_tool(tool_call["name"], tool_call["params"])
             tool_call["result"] = tool_result
-            
-            # If this is the speak tool, capture the final response
             if tool_call["name"] == "speak":
                 processing_result["final_response"] = tool_call["params"].get("text", "")
-                agent_logger.info(f"Speak tool executed with text: {processing_result['final_response']}")
-            else:
-                agent_logger.debug(f"Tool execution result: {tool_result}")
-                
             processing_result["tool_executions"].append(tool_call)
         except Exception as e:
             tool_call["error"] = str(e)
             processing_result["tool_executions"].append(tool_call)
             agent_logger.error(f"Error executing tool {tool_call['name']}: {e}")
             
-    def _parse_tool_call(self, line: str) -> Optional[Dict[str, Any]]:
-        """Parse a tool call from a line of text"""
-        import re
-        import json
-        
-        # First try to parse as JSON if it looks like JSON
-        line = line.strip()
-        if line.startswith('```json'):
+    def _parse_tool_calls(self, text: str) -> List[Dict[str, Any]]:
+        """Parse tool calls using ast.literal_eval for robustness."""
+        import ast
+        calls = []
+        text = text.strip()
+        if text.startswith("speak(") and text.endswith(")"):
             try:
-                # Extract JSON content
-                json_content = line[7:]  # Remove ```json
-                if json_content.endswith('```'):
-                    json_content = json_content[:-3]  # Remove trailing ```
-                json_content = json_content.strip()
-                
-                # Parse the JSON
-                tool_call_data = json.loads(json_content)
+                # Extract the content inside speak(...)
+                # e.g., "text='Hello, I'm here'"
+                inner_content = text[len("speak("):-1].strip()
+
+                # Ensure it's a text=... call
+                if not inner_content.startswith("text="):
+                    return []
                 
-                # Handle different JSON formats
-                if isinstance(tool_call_data, dict):
-                    # Check if it's a tool_code format
-                    if 'tool_code' in tool_call_data:
-                        # Extract the tool call from tool_code
-                        tool_code = tool_call_data['tool_code']
-                        # Remove any wrapper functions like print()
-                        tool_code = re.sub(r'^\w+\((.*)\)$', r'\1', tool_code)
-                        # Now parse the tool call normally
-                        pattern = r'(\w+)\((.*)\)'
-                        match = re.match(pattern, tool_code)
-                        if match:
-                            tool_name = match.group(1)
-                            params_str = match.group(2)
-                            
-                            # Parse parameters
-                            params = {}
-                            param_pattern = r'(\w+)\s*=\s*(".*?"|\'.*?\'|[^,]+?)(?:,|$)'
-                            for param_match in re.finditer(param_pattern, params_str):
-                                key, value = param_match.groups()
-                                # Remove quotes if present
-                                if (value.startswith('"') and value.endswith('"')) or \
-                                   (value.startswith("'") and value.endswith("'")):
-                                    value = value[1:-1]
-                                params[key] = value
-                                
-                            return {
-                                "name": tool_name,
-                                "params": params
-                            }
-                    # Check if it's a name/arguments format
-                    elif 'name' in tool_call_data and 'arguments' in tool_call_data:
-                        return {
-                            "name": tool_call_data['name'],
-                            "params": tool_call_data['arguments']
-                        }
-                elif isinstance(tool_call_data, list) and len(tool_call_data) > 0:
-                    # Handle array format - take the first item
-                    first_item = tool_call_data[0]
-                    if isinstance(first_item, dict):
-                        if 'tool_code' in first_item:
-                            # Extract the tool call from tool_code
-                            tool_code = first_item['tool_code']
-                            # Remove any wrapper functions like print()
-                            tool_code = re.sub(r'^\w+\((.*)\)$', r'\1', tool_code)
-                            # Now parse the tool call normally
-                            pattern = r'(\w+)\((.*)\)'
-                            match = re.match(pattern, tool_code)
-                            if match:
-                                tool_name = match.group(1)
-                                params_str = match.group(2)
-                                
-                                # Parse parameters
-                                params = {}
-                                param_pattern = r'(\w+)\s*=\s*(".*?"|\'.*?\'|[^,]+?)(?:,|$)'
-                                for param_match in re.finditer(param_pattern, params_str):
-                                    key, value = param_match.groups()
-                                    # Remove quotes if present
-                                    if (value.startswith('"') and value.endswith('"')) or \
-                                       (value.startswith("'") and value.endswith("'")):
-                                        value = value[1:-1]
-                                    params[key] = value
-                                    
-                                return {
-                                    "name": tool_name,
-                                    "params": params
-                                }
-                        elif 'name' in first_item and 'arguments' in first_item:
-                            return {
-                                "name": first_item['name'],
-                                "params": first_item['arguments']
-                            }
-                    
-            except (json.JSONDecodeError, KeyError, IndexError):
-                pass  # Fall back to regex parsing
-        
-        # Handle multi-line JSON that might be split across several lines
-        if line == '```json' or line == '{' or line == '}':
-            # Skip these lines as they're part of JSON structure
-            return None
-            
-        # Pattern to match tool_name(param1=value1, param2=value2, ...)
-        pattern = r'(\w+)\((.*)\)'
-        match = re.match(pattern, line)
-        
-        if match:
-            tool_name = match.group(1)
-            params_str = match.group(2)
-            
-            # Parse parameters more robustly
-            params = {}
-            
-            # Handle parameters one by one
-            # This handles quoted strings correctly, including special characters
-            param_pattern = r'(\w+)\s*=\s*(".*?"|\'.*?\'|[^,]+?)(?:,|$)'
-            for param_match in re.finditer(param_pattern, params_str):
-                key, value = param_match.groups()
-                # Remove quotes if present
-                if (value.startswith('"') and value.endswith('"')) or \
-                   (value.startswith("'") and value.endswith("'")):
-                    value = value[1:-1]
-                params[key] = value
+                # Get the quoted string part
+                quoted_string = inner_content[len("text="):
+].strip()
+
+                # Use ast.literal_eval to safely parse the Python string literal
+                parsed_text = ast.literal_eval(quoted_string)
                 
-            return {
-                "name": tool_name,
-                "params": params
-            }
-        
-        return None
+                if isinstance(parsed_text, str):
+                    calls.append({
+                        "name": "speak",
+                        "params": {"text": parsed_text}
+                    })
+
+            except (ValueError, SyntaxError, TypeError) as e:
+                agent_logger.warning(f"Could not parse tool call using ast.literal_eval: {text}. Error: {e}")
+
+        return calls
         
     async def execute_tool(self, tool_name: str, params: Dict[str, Any]) -> Any:
-        """Execute a registered tool"""
-        # Ensure agent is initialized
+        """Execute a registered tool."""
         await self.initialize()
         agent_logger.debug(f"Executing tool: {tool_name} with params: {params}")
         result = await self.tool_manager.execute_tool(tool_name, params)
         agent_logger.debug(f"Tool execution result: {result}")
         return result
-
-# Function to get agent logs (now uses the shared queue)
-def get_agent_logs(lines: int = 50) -> List[str]:
-    """Get recent agent logs from the shared queue"""
-    logs_list = list(agent_log_queue)
-    return logs_list[-lines:] if len(logs_list) > lines else logs_list

neuro_simulator/agent/factory.py

@@ -0,0 +1,30 @@
+# agent/factory.py
+"""Factory for creating agent instances"""
+
+from .base import BaseAgent
+from ..config import config_manager
+
+
+async def create_agent() -> BaseAgent:
+    """Create an agent instance based on the configuration"""
+    agent_type = config_manager.settings.agent_type
+    
+    if agent_type == "builtin":
+        from ..builtin_agent import local_agent, BuiltinAgentWrapper, initialize_builtin_agent
+        if local_agent is None:
+            # Try to initialize the builtin agent
+            await initialize_builtin_agent()
+            # Re-import local_agent after initialization
+            from ..builtin_agent import local_agent
+            if local_agent is None:
+                raise RuntimeError("Failed to initialize Builtin agent")
+        return BuiltinAgentWrapper(local_agent)
+    elif agent_type == "letta":
+        from ..letta import get_letta_agent, initialize_letta_client
+        # Try to initialize the letta client
+        initialize_letta_client()
+        agent = get_letta_agent()
+        await agent.initialize()
+        return agent
+    else:
+        raise ValueError(f"Unknown agent type: {agent_type}")