praisonaiagents 0.0.122__py3-none-any.whl → 0.0.124__py3-none-any.whl

praisonaiagents/agent/agent.py

@@ -444,6 +444,8 @@ class Agent:
         self.embedder_config = embedder_config
         self.knowledge = knowledge
         self.use_system_prompt = use_system_prompt
+        # NOTE: chat_history is not thread-safe. If concurrent access is needed,
+        # consider using threading.Lock or other synchronization mechanisms
         self.chat_history = []
         self.markdown = markdown
         self.stream = stream
@@ -711,8 +713,55 @@ Your Goal: {self.goal}
                 )
         
         return current_response
+    
+    def _build_system_prompt(self, tools=None):
+        """Build the system prompt with tool information.
+        
+        Args:
+            tools: Optional list of tools to use (defaults to self.tools)
+            
+        Returns:
+            str: The system prompt or None if use_system_prompt is False
+        """
+        if not self.use_system_prompt:
+            return None
+            
+        system_prompt = f"""{self.backstory}\n
+Your Role: {self.role}\n
+Your Goal: {self.goal}"""
+        
+        # Add tool usage instructions if tools are available
+        # Use provided tools or fall back to self.tools
+        tools_to_use = tools if tools is not None else self.tools
+        if tools_to_use:
+            tool_names = []
+            for tool in tools_to_use:
+                try:
+                    if callable(tool) and hasattr(tool, '__name__'):
+                        tool_names.append(tool.__name__)
+                    elif isinstance(tool, dict) and isinstance(tool.get('function'), dict) and 'name' in tool['function']:
+                        tool_names.append(tool['function']['name'])
+                    elif isinstance(tool, str):
+                        tool_names.append(tool)
+                    elif hasattr(tool, "to_openai_tool"):
+                        # Handle MCP tools
+                        openai_tools = tool.to_openai_tool()
+                        if isinstance(openai_tools, list):
+                            for t in openai_tools:
+                                if isinstance(t, dict) and 'function' in t and 'name' in t['function']:
+                                    tool_names.append(t['function']['name'])
+                        elif isinstance(openai_tools, dict) and 'function' in openai_tools:
+                            tool_names.append(openai_tools['function']['name'])
+                except (AttributeError, KeyError, TypeError) as e:
+                    logging.warning(f"Could not extract tool name from {tool}: {e}")
+                    continue
+            
+            if tool_names:
+                system_prompt += f"\n\nYou have access to the following tools: {', '.join(tool_names)}. Use these tools when appropriate to help complete your tasks. Always use tools when they can help provide accurate information or perform actions."
+        
+        return system_prompt
 
-    def _build_messages(self, prompt, temperature=0.2, output_json=None, output_pydantic=None):
+    def _build_messages(self, prompt, temperature=0.2, output_json=None, output_pydantic=None, tools=None):
         """Build messages list for chat completion.
         
         Args:
@@ -720,17 +769,13 @@ Your Goal: {self.goal}
             temperature: Temperature for the chat
             output_json: Optional Pydantic model for JSON output
             output_pydantic: Optional Pydantic model for JSON output (alias)
+            tools: Optional list of tools to use (defaults to self.tools)
             
         Returns:
             tuple: (messages list, original prompt)
         """
-        # Build system prompt if enabled
-        system_prompt = None
-        if self.use_system_prompt:
-            system_prompt = f"""{self.backstory}\n
-Your Role: {self.role}\n
-Your Goal: {self.goal}
-            """
+        # Build system prompt using the helper method
+        system_prompt = self._build_system_prompt(tools)
         
         # Use openai_client's build_messages method if available
         if self._openai_client is not None:
@@ -1154,10 +1199,27 @@ Your Goal: {self.goal}
                                 tool_param = [openai_tool]
                             logging.debug(f"Converted MCP tool: {tool_param}")
                 
-                # Pass everything to LLM class
-                response_text = self.llm_instance.get_response(
+                # Store chat history length for potential rollback
+                chat_history_length = len(self.chat_history)
+                
+                # Normalize prompt content for consistent chat history storage
+                normalized_content = prompt
+                if isinstance(prompt, list):
+                    # Extract text from multimodal prompts
+                    normalized_content = next((item["text"] for item in prompt if item.get("type") == "text"), "")
+                
+                # Prevent duplicate messages
+                if not (self.chat_history and 
+                        self.chat_history[-1].get("role") == "user" and 
+                        self.chat_history[-1].get("content") == normalized_content):
+                    # Add user message to chat history BEFORE LLM call so handoffs can access it
+                    self.chat_history.append({"role": "user", "content": normalized_content})
+                
+                try:
+                    # Pass everything to LLM class
+                    response_text = self.llm_instance.get_response(
                     prompt=prompt,
-                    system_prompt=f"{self.backstory}\n\nYour Role: {self.role}\n\nYour Goal: {self.goal}" if self.use_system_prompt else None,
+                    system_prompt=self._build_system_prompt(tools),
                     chat_history=self.chat_history,
                     temperature=temperature,
                     tools=tool_param,
@@ -1174,22 +1236,28 @@ Your Goal: {self.goal}
                     agent_tools=[t.__name__ if hasattr(t, '__name__') else str(t) for t in (tools if tools is not None else self.tools)],
                     execute_tool_fn=self.execute_tool,  # Pass tool execution function
                     reasoning_steps=reasoning_steps
-                )
+                    )
 
-                self.chat_history.append({"role": "user", "content": prompt})
-                self.chat_history.append({"role": "assistant", "content": response_text})
+                    self.chat_history.append({"role": "assistant", "content": response_text})
 
-                # Log completion time if in debug mode
-                if logging.getLogger().getEffectiveLevel() == logging.DEBUG:
-                    total_time = time.time() - start_time
-                    logging.debug(f"Agent.chat completed in {total_time:.2f} seconds")
+                    # Log completion time if in debug mode
+                    if logging.getLogger().getEffectiveLevel() == logging.DEBUG:
+                        total_time = time.time() - start_time
+                        logging.debug(f"Agent.chat completed in {total_time:.2f} seconds")
 
-                # Apply guardrail validation for custom LLM response
-                try:
-                    validated_response = self._apply_guardrail_with_retry(response_text, prompt, temperature, tools)
-                    return validated_response
+                    # Apply guardrail validation for custom LLM response
+                    try:
+                        validated_response = self._apply_guardrail_with_retry(response_text, prompt, temperature, tools)
+                        return validated_response
+                    except Exception as e:
+                        logging.error(f"Agent {self.name}: Guardrail validation failed for custom LLM: {e}")
+                        # Rollback chat history on guardrail failure
+                        self.chat_history = self.chat_history[:chat_history_length]
+                        return None
                 except Exception as e:
-                    logging.error(f"Agent {self.name}: Guardrail validation failed for custom LLM: {e}")
+                    # Rollback chat history if LLM call fails
+                    self.chat_history = self.chat_history[:chat_history_length]
+                    display_error(f"Error in LLM chat: {e}")
                     return None
             except Exception as e:
                 display_error(f"Error in LLM chat: {e}")
@@ -1197,191 +1265,209 @@ Your Goal: {self.goal}
         else:
             # Use the new _build_messages helper method
             messages, original_prompt = self._build_messages(prompt, temperature, output_json, output_pydantic)
+            
+            # Store chat history length for potential rollback
+            chat_history_length = len(self.chat_history)
+            
+            # Normalize original_prompt for consistent chat history storage
+            normalized_content = original_prompt
+            if isinstance(original_prompt, list):
+                # Extract text from multimodal prompts
+                normalized_content = next((item["text"] for item in original_prompt if item.get("type") == "text"), "")
+            
+            # Prevent duplicate messages
+            if not (self.chat_history and 
+                    self.chat_history[-1].get("role") == "user" and 
+                    self.chat_history[-1].get("content") == normalized_content):
+                # Add user message to chat history BEFORE LLM call so handoffs can access it
+                self.chat_history.append({"role": "user", "content": normalized_content})
 
-            final_response_text = None
             reflection_count = 0
             start_time = time.time()
-
-            while True:
-                try:
-                    if self.verbose:
-                        # Handle both string and list prompts for instruction display
-                        display_text = prompt
-                        if isinstance(prompt, list):
-                            # Extract text content from multimodal prompt
-                            display_text = next((item["text"] for item in prompt if item["type"] == "text"), "")
-                        
-                        if display_text and str(display_text).strip():
-                            # Pass agent information to display_instruction
-                            agent_tools = [t.__name__ if hasattr(t, '__name__') else str(t) for t in self.tools]
-                            display_instruction(
-                                f"Agent {self.name} is processing prompt: {display_text}", 
-                                console=self.console,
-                                agent_name=self.name,
-                                agent_role=self.role,
-                                agent_tools=agent_tools
-                            )
-
-                    response = self._chat_completion(messages, temperature=temperature, tools=tools if tools else None, reasoning_steps=reasoning_steps, stream=self.stream)
-                    if not response:
-                        return None
-
-                    response_text = response.choices[0].message.content.strip()
-
-                    # Handle output_json or output_pydantic if specified
-                    if output_json or output_pydantic:
-                        # Add to chat history and return raw response
-                        self.chat_history.append({"role": "user", "content": original_prompt})
-                        self.chat_history.append({"role": "assistant", "content": response_text})
-                        # Only display interaction if not using custom LLM (to avoid double output) and verbose is True
-                        if self.verbose and not self._using_custom_llm:
-                            display_interaction(original_prompt, response_text, markdown=self.markdown, 
-                                             generation_time=time.time() - start_time, console=self.console)
-                        return response_text
-
-                    if not self.self_reflect:
-                        self.chat_history.append({"role": "user", "content": original_prompt})
-                        self.chat_history.append({"role": "assistant", "content": response_text})
-                        if self.verbose:
-                            logging.debug(f"Agent {self.name} final response: {response_text}")
-                        # Only display interaction if not using custom LLM (to avoid double output) and verbose is True
-                        if self.verbose and not self._using_custom_llm:
-                            display_interaction(original_prompt, response_text, markdown=self.markdown, generation_time=time.time() - start_time, console=self.console)
-                        # Return only reasoning content if reasoning_steps is True
-                        if reasoning_steps and hasattr(response.choices[0].message, 'reasoning_content'):
-                            # Apply guardrail to reasoning content
-                            try:
-                                validated_reasoning = self._apply_guardrail_with_retry(response.choices[0].message.reasoning_content, original_prompt, temperature, tools)
-                                return validated_reasoning
-                            except Exception as e:
-                                logging.error(f"Agent {self.name}: Guardrail validation failed for reasoning content: {e}")
-                                return None
-                        # Apply guardrail to regular response
-                        try:
-                            validated_response = self._apply_guardrail_with_retry(response_text, original_prompt, temperature, tools)
-                            return validated_response
-                        except Exception as e:
-                            logging.error(f"Agent {self.name}: Guardrail validation failed: {e}")
-                            return None
-
-                    reflection_prompt = f"""
-Reflect on your previous response: '{response_text}'.
-{self.reflect_prompt if self.reflect_prompt else "Identify any flaws, improvements, or actions."}
-Provide a "satisfactory" status ('yes' or 'no').
-Output MUST be JSON with 'reflection' and 'satisfactory'.
-                    """
-                    logging.debug(f"{self.name} reflection attempt {reflection_count+1}, sending prompt: {reflection_prompt}")
-                    messages.append({"role": "user", "content": reflection_prompt})
-
+            
+            # Wrap entire while loop in try-except for rollback on any failure
+            try:
+                while True:
                     try:
-                        # Check if we're using a custom LLM (like Gemini)
-                        if self._using_custom_llm or self._openai_client is None:
-                            # For custom LLMs, we need to handle reflection differently
-                            # Use non-streaming to get complete JSON response
-                            reflection_response = self._chat_completion(messages, temperature=temperature, tools=None, stream=False, reasoning_steps=False)
-                            
-                            if not reflection_response or not reflection_response.choices:
-                                raise Exception("No response from reflection request")
-                            
-                            reflection_text = reflection_response.choices[0].message.content.strip()
-                            
-                            # Clean the JSON output
-                            cleaned_json = self.clean_json_output(reflection_text)
-                            
-                            # Parse the JSON manually
-                            reflection_data = json.loads(cleaned_json)
-                            
-                            # Create a reflection output object manually
-                            class CustomReflectionOutput:
-                                def __init__(self, data):
-                                    self.reflection = data.get('reflection', '')
-                                    self.satisfactory = data.get('satisfactory', 'no').lower()
-                            
-                            reflection_output = CustomReflectionOutput(reflection_data)
-                        else:
-                            # Use OpenAI's structured output for OpenAI models
-                            reflection_response = self._openai_client.sync_client.beta.chat.completions.parse(
-                                model=self.reflect_llm if self.reflect_llm else self.llm,
-                                messages=messages,
-                                temperature=temperature,
-                                response_format=ReflectionOutput
-                            )
-
-                            reflection_output = reflection_response.choices[0].message.parsed
-
                         if self.verbose:
-                            display_self_reflection(f"Agent {self.name} self reflection (using {self.reflect_llm if self.reflect_llm else self.llm}): reflection='{reflection_output.reflection}' satisfactory='{reflection_output.satisfactory}'", console=self.console)
+                            # Handle both string and list prompts for instruction display
+                            display_text = prompt
+                            if isinstance(prompt, list):
+                                # Extract text content from multimodal prompt
+                                display_text = next((item["text"] for item in prompt if item["type"] == "text"), "")
+                            
+                            if display_text and str(display_text).strip():
+                                # Pass agent information to display_instruction
+                                agent_tools = [t.__name__ if hasattr(t, '__name__') else str(t) for t in self.tools]
+                                display_instruction(
+                                    f"Agent {self.name} is processing prompt: {display_text}", 
+                                    console=self.console,
+                                    agent_name=self.name,
+                                    agent_role=self.role,
+                                    agent_tools=agent_tools
+                                )
+
+                        response = self._chat_completion(messages, temperature=temperature, tools=tools if tools else None, reasoning_steps=reasoning_steps, stream=self.stream)
+                        if not response:
+                            # Rollback chat history on response failure
+                            self.chat_history = self.chat_history[:chat_history_length]
+                            return None
 
-                        messages.append({"role": "assistant", "content": f"Self Reflection: {reflection_output.reflection} Satisfactory?: {reflection_output.satisfactory}"})
+                        response_text = response.choices[0].message.content.strip()
 
-                        # Only consider satisfactory after minimum reflections
-                        if reflection_output.satisfactory == "yes" and reflection_count >= self.min_reflect - 1:
-                            if self.verbose:
-                                display_self_reflection("Agent marked the response as satisfactory after meeting minimum reflections", console=self.console)
-                            self.chat_history.append({"role": "user", "content": prompt})
+                        # Handle output_json or output_pydantic if specified
+                        if output_json or output_pydantic:
+                            # Add to chat history and return raw response
+                            # User message already added before LLM call via _build_messages
                             self.chat_history.append({"role": "assistant", "content": response_text})
                             # Only display interaction if not using custom LLM (to avoid double output) and verbose is True
                             if self.verbose and not self._using_custom_llm:
-                                display_interaction(prompt, response_text, markdown=self.markdown, generation_time=time.time() - start_time, console=self.console)
-                            # Apply guardrail validation after satisfactory reflection
-                            try:
-                                validated_response = self._apply_guardrail_with_retry(response_text, prompt, temperature, tools)
-                                return validated_response
-                            except Exception as e:
-                                logging.error(f"Agent {self.name}: Guardrail validation failed after reflection: {e}")
-                                return None
+                                display_interaction(original_prompt, response_text, markdown=self.markdown, 
+                                                 generation_time=time.time() - start_time, console=self.console)
+                            return response_text
 
-                        # Check if we've hit max reflections
-                        if reflection_count >= self.max_reflect - 1:
-                            if self.verbose:
-                                display_self_reflection("Maximum reflection count reached, returning current response", console=self.console)
-                            self.chat_history.append({"role": "user", "content": prompt})
+                        if not self.self_reflect:
+                            # User message already added before LLM call via _build_messages
                             self.chat_history.append({"role": "assistant", "content": response_text})
+                            if self.verbose:
+                                logging.debug(f"Agent {self.name} final response: {response_text}")
                             # Only display interaction if not using custom LLM (to avoid double output) and verbose is True
                             if self.verbose and not self._using_custom_llm:
-                                display_interaction(prompt, response_text, markdown=self.markdown, generation_time=time.time() - start_time, console=self.console)
-                            # Apply guardrail validation after max reflections
+                                display_interaction(original_prompt, response_text, markdown=self.markdown, generation_time=time.time() - start_time, console=self.console)
+                            # Return only reasoning content if reasoning_steps is True
+                            if reasoning_steps and hasattr(response.choices[0].message, 'reasoning_content'):
+                                # Apply guardrail to reasoning content
+                                try:
+                                    validated_reasoning = self._apply_guardrail_with_retry(response.choices[0].message.reasoning_content, original_prompt, temperature, tools)
+                                    return validated_reasoning
+                                except Exception as e:
+                                    logging.error(f"Agent {self.name}: Guardrail validation failed for reasoning content: {e}")
+                                    # Rollback chat history on guardrail failure
+                                    self.chat_history = self.chat_history[:chat_history_length]
+                                    return None
+                            # Apply guardrail to regular response
                             try:
-                                validated_response = self._apply_guardrail_with_retry(response_text, prompt, temperature, tools)
+                                validated_response = self._apply_guardrail_with_retry(response_text, original_prompt, temperature, tools)
                                 return validated_response
                             except Exception as e:
-                                logging.error(f"Agent {self.name}: Guardrail validation failed after max reflections: {e}")
+                                logging.error(f"Agent {self.name}: Guardrail validation failed: {e}")
+                                # Rollback chat history on guardrail failure
+                                self.chat_history = self.chat_history[:chat_history_length]
                                 return None
 
-                        logging.debug(f"{self.name} reflection count {reflection_count + 1}, continuing reflection process")
-                        messages.append({"role": "user", "content": "Now regenerate your response using the reflection you made"})
-                        # For custom LLMs during reflection, always use non-streaming to ensure complete responses
-                        use_stream = self.stream if not self._using_custom_llm else False
-                        response = self._chat_completion(messages, temperature=temperature, tools=None, stream=use_stream)
-                        response_text = response.choices[0].message.content.strip()
-                        reflection_count += 1
-                        continue  # Continue the loop for more reflections
+                        reflection_prompt = f"""
+Reflect on your previous response: '{response_text}'.
+{self.reflect_prompt if self.reflect_prompt else "Identify any flaws, improvements, or actions."}
+Provide a "satisfactory" status ('yes' or 'no').
+Output MUST be JSON with 'reflection' and 'satisfactory'.
+                        """
+                        logging.debug(f"{self.name} reflection attempt {reflection_count+1}, sending prompt: {reflection_prompt}")
+                        messages.append({"role": "user", "content": reflection_prompt})
 
-                    except Exception as e:
-                        display_error(f"Error in parsing self-reflection json {e}. Retrying", console=self.console)
-                        logging.error("Reflection parsing failed.", exc_info=True)
-                        messages.append({"role": "assistant", "content": f"Self Reflection failed."})
-                        reflection_count += 1
-                        continue  # Continue even after error to try again
-                    
-                except Exception as e:
-                    display_error(f"Error in chat: {e}", console=self.console)
-                    return None 
+                        try:
+                            # Check if we're using a custom LLM (like Gemini)
+                            if self._using_custom_llm or self._openai_client is None:
+                                # For custom LLMs, we need to handle reflection differently
+                                # Use non-streaming to get complete JSON response
+                                reflection_response = self._chat_completion(messages, temperature=temperature, tools=None, stream=False, reasoning_steps=False)
+                                
+                                if not reflection_response or not reflection_response.choices:
+                                    raise Exception("No response from reflection request")
+                                
+                                reflection_text = reflection_response.choices[0].message.content.strip()
+                                
+                                # Clean the JSON output
+                                cleaned_json = self.clean_json_output(reflection_text)
+                                
+                                # Parse the JSON manually
+                                reflection_data = json.loads(cleaned_json)
+                                
+                                # Create a reflection output object manually
+                                class CustomReflectionOutput:
+                                    def __init__(self, data):
+                                        self.reflection = data.get('reflection', '')
+                                        self.satisfactory = data.get('satisfactory', 'no').lower()
+                                
+                                reflection_output = CustomReflectionOutput(reflection_data)
+                            else:
+                                # Use OpenAI's structured output for OpenAI models
+                                reflection_response = self._openai_client.sync_client.beta.chat.completions.parse(
+                                    model=self.reflect_llm if self.reflect_llm else self.llm,
+                                    messages=messages,
+                                    temperature=temperature,
+                                    response_format=ReflectionOutput
+                                )
 
-        # Log completion time if in debug mode
-        if logging.getLogger().getEffectiveLevel() == logging.DEBUG:
-            total_time = time.time() - start_time
-            logging.debug(f"Agent.chat completed in {total_time:.2f} seconds")
-        
-        # Apply guardrail validation before returning    
-        try:
-            validated_response = self._apply_guardrail_with_retry(response_text, prompt, temperature, tools)
-            return validated_response
-        except Exception as e:
-            logging.error(f"Agent {self.name}: Guardrail validation failed: {e}")
-            if self.verbose:
-                display_error(f"Guardrail validation failed: {e}", console=self.console)
-            return None
+                                reflection_output = reflection_response.choices[0].message.parsed
+
+                            if self.verbose:
+                                display_self_reflection(f"Agent {self.name} self reflection (using {self.reflect_llm if self.reflect_llm else self.llm}): reflection='{reflection_output.reflection}' satisfactory='{reflection_output.satisfactory}'", console=self.console)
+
+                            messages.append({"role": "assistant", "content": f"Self Reflection: {reflection_output.reflection} Satisfactory?: {reflection_output.satisfactory}"})
+
+                            # Only consider satisfactory after minimum reflections
+                            if reflection_output.satisfactory == "yes" and reflection_count >= self.min_reflect - 1:
+                                if self.verbose:
+                                    display_self_reflection("Agent marked the response as satisfactory after meeting minimum reflections", console=self.console)
+                                # User message already added before LLM call via _build_messages
+                                self.chat_history.append({"role": "assistant", "content": response_text})
+                                # Only display interaction if not using custom LLM (to avoid double output) and verbose is True
+                                if self.verbose and not self._using_custom_llm:
+                                    display_interaction(original_prompt, response_text, markdown=self.markdown, generation_time=time.time() - start_time, console=self.console)
+                                # Apply guardrail validation after satisfactory reflection
+                                try:
+                                    validated_response = self._apply_guardrail_with_retry(response_text, original_prompt, temperature, tools)
+                                    return validated_response
+                                except Exception as e:
+                                    logging.error(f"Agent {self.name}: Guardrail validation failed after reflection: {e}")
+                                    # Rollback chat history on guardrail failure
+                                    self.chat_history = self.chat_history[:chat_history_length]
+                                    return None
+
+                            # Check if we've hit max reflections
+                            if reflection_count >= self.max_reflect - 1:
+                                if self.verbose:
+                                    display_self_reflection("Maximum reflection count reached, returning current response", console=self.console)
+                                # User message already added before LLM call via _build_messages
+                                self.chat_history.append({"role": "assistant", "content": response_text})
+                                # Only display interaction if not using custom LLM (to avoid double output) and verbose is True
+                                if self.verbose and not self._using_custom_llm:
+                                    display_interaction(original_prompt, response_text, markdown=self.markdown, generation_time=time.time() - start_time, console=self.console)
+                                # Apply guardrail validation after max reflections
+                                try:
+                                    validated_response = self._apply_guardrail_with_retry(response_text, original_prompt, temperature, tools)
+                                    return validated_response
+                                except Exception as e:
+                                    logging.error(f"Agent {self.name}: Guardrail validation failed after max reflections: {e}")
+                                    # Rollback chat history on guardrail failure
+                                    self.chat_history = self.chat_history[:chat_history_length]
+                                    return None
+                            
+                            # If not satisfactory and not at max reflections, continue with regeneration
+                            logging.debug(f"{self.name} reflection count {reflection_count + 1}, continuing reflection process")
+                            messages.append({"role": "user", "content": "Now regenerate your response using the reflection you made"})
+                            # For custom LLMs during reflection, always use non-streaming to ensure complete responses
+                            use_stream = self.stream if not self._using_custom_llm else False
+                            response = self._chat_completion(messages, temperature=temperature, tools=None, stream=use_stream)
+                            response_text = response.choices[0].message.content.strip()
+                            reflection_count += 1
+                            continue  # Continue the loop for more reflections
+
+                        except Exception as e:
+                                display_error(f"Error in parsing self-reflection json {e}. Retrying", console=self.console)
+                                logging.error("Reflection parsing failed.", exc_info=True)
+                                messages.append({"role": "assistant", "content": "Self Reflection failed."})
+                                reflection_count += 1
+                                continue  # Continue even after error to try again
+                    except Exception:
+                        # Catch any exception from the inner try block and re-raise to outer handler
+                        raise
+            except Exception as e:
+                # Catch any exceptions that escape the while loop
+                display_error(f"Unexpected error in chat: {e}", console=self.console)
+                # Rollback chat history
+                self.chat_history = self.chat_history[:chat_history_length]
+                return None
 
     def clean_json_output(self, output: str) -> str:
         """Clean and extract JSON from response text."""
@@ -1430,10 +1516,26 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                     prompt = f"{prompt}\n\nKnowledge: {knowledge_content}"
 
             if self._using_custom_llm:
+                # Store chat history length for potential rollback
+                chat_history_length = len(self.chat_history)
+                
+                # Normalize prompt content for consistent chat history storage
+                normalized_content = prompt
+                if isinstance(prompt, list):
+                    # Extract text from multimodal prompts
+                    normalized_content = next((item["text"] for item in prompt if item.get("type") == "text"), "")
+                
+                # Prevent duplicate messages
+                if not (self.chat_history and 
+                        self.chat_history[-1].get("role") == "user" and 
+                        self.chat_history[-1].get("content") == normalized_content):
+                    # Add user message to chat history BEFORE LLM call so handoffs can access it
+                    self.chat_history.append({"role": "user", "content": normalized_content})
+                
                 try:
                     response_text = await self.llm_instance.get_response_async(
                         prompt=prompt,
-                        system_prompt=f"{self.backstory}\n\nYour Role: {self.role}\n\nYour Goal: {self.goal}" if self.use_system_prompt else None,
+                        system_prompt=self._build_system_prompt(tools),
                         chat_history=self.chat_history,
                         temperature=temperature,
                         tools=tools,
@@ -1447,19 +1549,29 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                         console=self.console,
                         agent_name=self.name,
                         agent_role=self.role,
-                        agent_tools=[t.__name__ if hasattr(t, '__name__') else str(t) for t in self.tools],
+                        agent_tools=[t.__name__ if hasattr(t, '__name__') else str(t) for t in (tools if tools is not None else self.tools)],
                         execute_tool_fn=self.execute_tool_async,
                         reasoning_steps=reasoning_steps
                     )
 
-                    self.chat_history.append({"role": "user", "content": prompt})
                     self.chat_history.append({"role": "assistant", "content": response_text})
 
                     if logging.getLogger().getEffectiveLevel() == logging.DEBUG:
                         total_time = time.time() - start_time
                         logging.debug(f"Agent.achat completed in {total_time:.2f} seconds")
-                    return response_text
+                    
+                    # Apply guardrail validation for custom LLM response
+                    try:
+                        validated_response = self._apply_guardrail_with_retry(response_text, prompt, temperature, tools)
+                        return validated_response
+                    except Exception as e:
+                        logging.error(f"Agent {self.name}: Guardrail validation failed for custom LLM: {e}")
+                        # Rollback chat history on guardrail failure
+                        self.chat_history = self.chat_history[:chat_history_length]
+                        return None
                 except Exception as e:
+                    # Rollback chat history if LLM call fails
+                    self.chat_history = self.chat_history[:chat_history_length]
                     display_error(f"Error in LLM chat: {e}")
                     if logging.getLogger().getEffectiveLevel() == logging.DEBUG:
                         total_time = time.time() - start_time
@@ -1469,6 +1581,22 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
             # For OpenAI client
             # Use the new _build_messages helper method
             messages, original_prompt = self._build_messages(prompt, temperature, output_json, output_pydantic)
+            
+            # Store chat history length for potential rollback
+            chat_history_length = len(self.chat_history)
+            
+            # Normalize original_prompt for consistent chat history storage
+            normalized_content = original_prompt
+            if isinstance(original_prompt, list):
+                # Extract text from multimodal prompts
+                normalized_content = next((item["text"] for item in original_prompt if item.get("type") == "text"), "")
+            
+            # Prevent duplicate messages
+            if not (self.chat_history and 
+                    self.chat_history[-1].get("role") == "user" and 
+                    self.chat_history[-1].get("content") == normalized_content):
+                # Add user message to chat history BEFORE LLM call so handoffs can access it
+                self.chat_history.append({"role": "user", "content": normalized_content})
 
             reflection_count = 0
             start_time = time.time()
@@ -1615,7 +1743,16 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                         if logging.getLogger().getEffectiveLevel() == logging.DEBUG:
                             total_time = time.time() - start_time
                             logging.debug(f"Agent.achat completed in {total_time:.2f} seconds")
-                        return response_text
+                        
+                        # Apply guardrail validation for OpenAI client response
+                        try:
+                            validated_response = self._apply_guardrail_with_retry(response_text, original_prompt, temperature, tools)
+                            return validated_response
+                        except Exception as e:
+                            logging.error(f"Agent {self.name}: Guardrail validation failed for OpenAI client: {e}")
+                            # Rollback chat history on guardrail failure
+                            self.chat_history = self.chat_history[:chat_history_length]
+                            return None
                 except Exception as e:
                     display_error(f"Error in chat completion: {e}")
                     if logging.getLogger().getEffectiveLevel() == logging.DEBUG:

praisonaiagents/agents/autoagents.py

@@ -296,6 +296,10 @@ IMPORTANT: Each task MUST be an object with name, description, expected_output,
         last_error = None
         
         for attempt in range(max_retries):
+            # Initialize variables for this attempt
+            use_openai_structured = False
+            client = None
+            
             # Prepare prompt for this attempt
             if attempt > 0 and last_response and last_error:
                 # On retry, include the previous response and error
@@ -322,23 +326,16 @@ DO NOT use strings for tasks. Each task MUST be a complete object with all four
                 prompt = base_prompt
             
             try:
-                # Try to use OpenAI's structured output if available
-                use_openai_structured = False
-                client = None
-                
-                try:
-                    # Check if we have OpenAI API and the model supports structured output
-                    if self.llm and (self.llm.startswith('gpt-') or self.llm.startswith('o1-') or self.llm.startswith('o3-')):
-                        # Create a new client instance if custom parameters are provided
-                        if self.api_key or self.base_url:
-                            client = OpenAIClient(api_key=self.api_key, base_url=self.base_url)
-                        else:
-                            client = get_openai_client()
-                        use_openai_structured = True
-                except:
-                    # If OpenAI client is not available, we'll use the LLM class
-                    pass
-                
+                # Check if we have OpenAI API and the model supports structured output
+                from ..llm import supports_structured_outputs
+                if self.llm and supports_structured_outputs(self.llm):
+                    client = get_openai_client()
+                    use_openai_structured = True
+            except:
+                # If OpenAI client is not available, we'll use the LLM class
+                pass
+            
+            try:
                 if use_openai_structured and client:
                     # Use OpenAI's structured output for OpenAI models (backward compatibility)
                     config = client.parse_structured_output(
@@ -359,7 +356,7 @@ DO NOT use strings for tasks. Each task MUST be a complete object with all four
                         api_key=self.api_key
                     )
                     
-                    response_text = llm_instance.response(
+                    response_text = llm_instance.get_response(
                         prompt=prompt,
                         system_prompt="You are a helpful assistant designed to generate AI agent configurations.",
                         output_pydantic=AutoAgentsConfig,
@@ -503,4 +500,4 @@ DO NOT use strings for tasks. Each task MUST be a complete object with all four
         Creates tasks based on the instructions, then starts execution.
         Returns the task status and results dictionary.
         """
-        return super().start() 
+        return super().start()

praisonaiagents/llm/__init__.py

@@ -32,6 +32,10 @@ from .openai_client import (
     ToolCall,
     process_stream_chunks
 )
+from .model_capabilities import (
+    supports_structured_outputs,
+    supports_streaming_with_tools
+)
 
 # Ensure telemetry is disabled after import as well
 try:
@@ -52,5 +56,7 @@ __all__ = [
     "CompletionUsage",
     "ChatCompletion",
     "ToolCall",
-    "process_stream_chunks"
+    "process_stream_chunks",
+    "supports_structured_outputs",
+    "supports_streaming_with_tools"
 ]

praisonaiagents/llm/llm.py

@@ -386,8 +386,11 @@ class LLM:
         if self._is_ollama_provider():
             return False
         
-        # OpenAI models support streaming with tools
-        if any(self.model.startswith(prefix) for prefix in ["gpt-", "o1-", "o3-"]):
+        # Import the capability check function
+        from .model_capabilities import supports_streaming_with_tools
+        
+        # Check if this model supports streaming with tools
+        if supports_streaming_with_tools(self.model):
             return True
         
         # Anthropic Claude models support streaming with tools
@@ -403,7 +406,7 @@ class LLM:
         # missing tool calls or making duplicate calls
         return False
     
-    def _build_messages(self, prompt, system_prompt=None, chat_history=None, output_json=None, output_pydantic=None):
+    def _build_messages(self, prompt, system_prompt=None, chat_history=None, output_json=None, output_pydantic=None, tools=None):
         """Build messages list for LLM completion. Works for both sync and async.
         
         Args:
@@ -412,6 +415,7 @@ class LLM:
             chat_history: Optional list of previous messages
             output_json: Optional Pydantic model for JSON output
             output_pydantic: Optional Pydantic model for JSON output (alias)
+            tools: Optional list of tools available
             
         Returns:
             tuple: (messages list, original prompt)
@@ -1855,6 +1859,21 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
         # Override with any provided parameters
         params.update(override_params)
         
+        # Add tool_choice="auto" when tools are provided (unless already specified)
+        if 'tools' in params and params['tools'] and 'tool_choice' not in params:
+            # For Gemini models, use tool_choice to encourage tool usage
+            # More comprehensive Gemini model detection
+            if any(prefix in self.model.lower() for prefix in ['gemini', 'gemini/', 'google/gemini']):
+                try:
+                    import litellm
+                    # Check if model supports function calling before setting tool_choice
+                    if litellm.supports_function_calling(model=self.model):
+                        params['tool_choice'] = 'auto'
+                except Exception as e:
+                    # If check fails, still set tool_choice for known Gemini models
+                    logging.debug(f"Could not verify function calling support: {e}. Setting tool_choice anyway.")
+                    params['tool_choice'] = 'auto'
+        
         return params
 
     def _prepare_response_logging(self, temperature: float, stream: bool, verbose: bool, markdown: bool, **kwargs) -> Optional[Dict[str, Any]]:

praisonaiagents/llm/model_capabilities.py

@@ -0,0 +1,90 @@
+"""
+Model capabilities configuration for different LLM providers.
+This module defines which models support specific features like structured outputs.
+"""
+
+# Models that support OpenAI-style structured outputs (response_format with Pydantic models)
+MODELS_SUPPORTING_STRUCTURED_OUTPUTS = {
+    # OpenAI models
+    "gpt-4o",
+    "gpt-4o-mini",
+    "gpt-4-turbo",
+    "gpt-4-turbo-preview",
+    "gpt-4-turbo-2024-04-09",
+    "gpt-4-1106-preview",
+    "gpt-4-0125-preview",
+    "gpt-3.5-turbo",
+    "gpt-3.5-turbo-1106",
+    "gpt-3.5-turbo-0125",
+    
+    # New/Future OpenAI models (as mentioned by user)
+    "codex-mini",
+    "o3-pro",
+    "gpt-4.5-preview",
+    "o3-mini",
+    "o1",
+    "o1-preview",
+    "o1-mini",
+    "gpt-4.1",
+    "gpt-4.1-nano",
+    "gpt-4.1-mini",
+    "o4-mini",
+    "o3",
+}
+
+# Models that explicitly DON'T support structured outputs
+MODELS_NOT_SUPPORTING_STRUCTURED_OUTPUTS = {
+    # Audio preview models
+    "gpt-4o-audio-preview",
+    "gpt-4o-mini-audio-preview",
+    
+    # Legacy o1 models (don't support system messages either)
+    "o1-preview-2024-09-12",
+    "o1-mini-2024-09-12",
+}
+
+
+def supports_structured_outputs(model_name: str) -> bool:
+    """
+    Check if a model supports OpenAI-style structured outputs.
+    
+    Args:
+        model_name: The name of the model to check
+        
+    Returns:
+        bool: True if the model supports structured outputs, False otherwise
+    """
+    if not model_name:
+        return False
+    
+    # First check if it's explicitly in the NOT supporting list
+    if model_name in MODELS_NOT_SUPPORTING_STRUCTURED_OUTPUTS:
+        return False
+    
+    # Then check if it's in the supporting list
+    if model_name in MODELS_SUPPORTING_STRUCTURED_OUTPUTS:
+        return True
+    
+    # For models with version suffixes, check the base model name
+    base_model = model_name.split('-2024-')[0].split('-2025-')[0]
+    if base_model in MODELS_SUPPORTING_STRUCTURED_OUTPUTS:
+        return True
+    
+    # Default to False for unknown models
+    return False
+
+
+def supports_streaming_with_tools(model_name: str) -> bool:
+    """
+    Check if a model supports streaming when tools are provided.
+    Most models that support structured outputs also support streaming with tools.
+    
+    Args:
+        model_name: The name of the model to check
+        
+    Returns:
+        bool: True if the model supports streaming with tools, False otherwise
+    """
+    # For now, use the same logic as structured outputs
+    # In the future, this could be a separate list if needed
+    return supports_structured_outputs(model_name)

{praisonaiagents-0.0.122.dist-info → praisonaiagents-0.0.124.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: praisonaiagents
-Version: 0.0.122
+Version: 0.0.124
 Summary: Praison AI agents for completing complex tasks with Self Reflection Agents
 Author: Mervin Praison
 Requires-Python: >=3.10

{praisonaiagents-0.0.122.dist-info → praisonaiagents-0.0.124.dist-info}/RECORD

@@ -3,20 +3,21 @@ praisonaiagents/approval.py,sha256=UJ4OhfihpFGR5CAaMphqpSvqdZCHi5w2MGw1MByZ1FQ,9
 praisonaiagents/main.py,sha256=bamnEu5PaekloGi52VqAFclm-HzjEVeKtWF0Zpdmfzs,15479
 praisonaiagents/session.py,sha256=d-CZPYikOHb0q-H9f_IWKJsypnQfz1YKeLLkyxs6oDo,15532
 praisonaiagents/agent/__init__.py,sha256=IhIDtAkfJ99cxbttwou52coih_AejS2-jpazsX6LbDY,350
-praisonaiagents/agent/agent.py,sha256=BZx0iCP4hHfKFlfGlkZtanBJDDBDZe54dhf4Oskhmhw,110427
+praisonaiagents/agent/agent.py,sha256=9JB46swjvrPSm7pQnsawGd2mV50qA-tx3Vv05_OoJeU,119579
 praisonaiagents/agent/handoff.py,sha256=Saq0chqfvC6Zf5UbXvmctybbehqnotrXn72JsS-76Q0,13099
 praisonaiagents/agent/image_agent.py,sha256=-5MXG594HVwSpFMcidt16YBp7udtik-Cp7eXlzLE1fY,8696
 praisonaiagents/agents/__init__.py,sha256=_1d6Pqyk9EoBSo7E68sKyd1jDRlN1vxvVIRpoMc0Jcw,168
 praisonaiagents/agents/agents.py,sha256=WnptTEMSDMAM30Ka6rOAu6rBD-ZLev3qphb1a3BbP1g,63301
-praisonaiagents/agents/autoagents.py,sha256=1stF8z94eyVg6hyfrLgdArlarftz_OFvEDtRMsGZFvg,21094
+praisonaiagents/agents/autoagents.py,sha256=BIbsl1m7SzZOh5BO2dfV_xi8sivGSWU9Xj4lhAkFru8,20765
 praisonaiagents/guardrails/__init__.py,sha256=HA8zhp-KRHTxo0194MUwXOUJjPyjOu7E3d7xUIKYVVY,310
 praisonaiagents/guardrails/guardrail_result.py,sha256=2K1WIYRyT_s1H6vBGa-7HEHzXCFIyZXZVY4f0hnQyWc,1352
 praisonaiagents/guardrails/llm_guardrail.py,sha256=MTTqmYDdZX-18QN9T17T5P_6H2qnV8GVgymJufW1WuM,3277
 praisonaiagents/knowledge/__init__.py,sha256=xL1Eh-a3xsHyIcU4foOWF-JdWYIYBALJH9bge0Ujuto,246
 praisonaiagents/knowledge/chunking.py,sha256=G6wyHa7_8V0_7VpnrrUXbEmUmptlT16ISJYaxmkSgmU,7678
 praisonaiagents/knowledge/knowledge.py,sha256=OKPar-XGyAp1ndmbOOdCgqFnTCqpOThYVSIZRxZyP58,15683
-praisonaiagents/llm/__init__.py,sha256=6lTeQ8jWi1-KiwjCDCmkHo2e-bRLq2dP0s5iJWqjO3s,1421
-praisonaiagents/llm/llm.py,sha256=8cDahPVMPI882J0psA1cXreJGXvO33eSOpMNy7FLCS4,107383
+praisonaiagents/llm/__init__.py,sha256=2rCx-Vfgg881XhVnOSDIt8mDab3Nj_BkHwDP4Bbbudc,1596
+praisonaiagents/llm/llm.py,sha256=9AQdZzd12d_DY6A_jPPfFSHhiEYaqZidVkguoY8j8Es,108522
+praisonaiagents/llm/model_capabilities.py,sha256=poxOxATUOi9XPTx3v6BPnXvSfikWSA9NciWQVuPU7Zg,2586
 praisonaiagents/llm/openai_client.py,sha256=0JvjCDHoH8I8kIt5vvObARkGdVaPWdTIv_FoEQ5EQPA,48973
 praisonaiagents/mcp/__init__.py,sha256=ibbqe3_7XB7VrIcUcetkZiUZS1fTVvyMy_AqCSFG8qc,240
 praisonaiagents/mcp/mcp.py,sha256=-fFx4MHffnN2woLnnV7Pzx3-1SFkn2j8Gp5F5ZIwKJ0,19698
@@ -53,7 +54,7 @@ praisonaiagents/tools/xml_tools.py,sha256=iYTMBEk5l3L3ryQ1fkUnNVYK-Nnua2Kx2S0dxN
 praisonaiagents/tools/yaml_tools.py,sha256=uogAZrhXV9O7xvspAtcTfpKSQYL2nlOTvCQXN94-G9A,14215
 praisonaiagents/tools/yfinance_tools.py,sha256=s2PBj_1v7oQnOobo2fDbQBACEHl61ftG4beG6Z979ZE,8529
 praisonaiagents/tools/train/data/generatecot.py,sha256=H6bNh-E2hqL5MW6kX3hqZ05g9ETKN2-kudSjiuU_SD8,19403
-praisonaiagents-0.0.122.dist-info/METADATA,sha256=pU6W0akH1O1raC15FOsFQW3GXuflwpznV2ij10vYUP4,1669
-praisonaiagents-0.0.122.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-praisonaiagents-0.0.122.dist-info/top_level.txt,sha256=_HsRddrJ23iDx5TTqVUVvXG2HeHBL5voshncAMDGjtA,16
-praisonaiagents-0.0.122.dist-info/RECORD,,
+praisonaiagents-0.0.124.dist-info/METADATA,sha256=_yFtjXGrQjKln7yCjLdflt8r0CGBVCpy-8EBFDzeMJQ,1669
+praisonaiagents-0.0.124.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+praisonaiagents-0.0.124.dist-info/top_level.txt,sha256=_HsRddrJ23iDx5TTqVUVvXG2HeHBL5voshncAMDGjtA,16
+praisonaiagents-0.0.124.dist-info/RECORD,,