ag2 0.9.3__py3-none-any.whl → 0.9.5__py3-none-any.whl

autogen/agentchat/groupchat.py

@@ -1023,6 +1023,38 @@ class GroupChat:
                 mentions[agent.name] = count
         return mentions
 
+    def _run_input_guardrails(
+        self,
+        agent: "ConversableAgent",
+        messages: Optional[list[dict[str, Any]]] = None,
+    ) -> Optional[str]:
+        """Run input guardrails for an agent before the reply is generated.
+        Args:
+            agent (ConversableAgent): The agent whose input guardrails to run.
+            messages (Optional[list[dict[str, Any]]]): The messages to check against the guardrails.
+        """
+        for guardrail in agent.input_guardrails:
+            guardrail_result = guardrail.check(context=messages)
+
+            if guardrail_result.activated:
+                guardrail.target.activate_target(self)
+                return f"{guardrail.activation_message}\nJustification: {guardrail_result.justification}"
+        return None
+
+    def _run_output_guardrails(self, agent: "ConversableAgent", reply: str) -> None:
+        """Run output guardrails for an agent after the reply is generated.
+        Args:
+            agent (ConversableAgent): The agent whose output guardrails to run.
+            reply (str): The reply generated by the agent.
+        """
+        for guardrail in agent.output_guardrails:
+            guardrail_result = guardrail.check(context=reply)
+
+            if guardrail_result.activated:
+                guardrail.target.activate_target(self)
+                return f"{guardrail.activation_message}\nJustification: {guardrail_result.justification}"
+        return None
+
 
 @export_module("autogen")
 class GroupChatManager(ConversableAgent):
@@ -1194,8 +1226,17 @@ class GroupChatManager(ConversableAgent):
                 if not silent:
                     iostream = IOStream.get_default()
                     iostream.send(GroupChatRunChatEvent(speaker=speaker, silent=silent))
-                # let the speaker speak
-                reply = speaker.generate_reply(sender=self)
+
+                guardrails_activated = False
+                guardrails_reply = groupchat._run_input_guardrails(speaker, speaker._oai_messages[self])
+
+                if guardrails_reply is not None:
+                    # if a guardrail has been activated, then the next target has been set and the guardrail reply will be sent
+                    guardrails_activated = True
+                    reply = guardrails_reply
+                else:
+                    # let the speaker speak
+                    reply = speaker.generate_reply(sender=self)
             except KeyboardInterrupt:
                 # let the admin agent speak if interrupted
                 if groupchat.admin_name in groupchat.agent_names:
@@ -1215,6 +1256,15 @@ class GroupChatManager(ConversableAgent):
                 termination_reason = "No reply generated"
                 break
 
+            if not guardrails_activated:
+                # if the input guardrails were not activated, and the agent returned a reply
+                guardrails_reply = groupchat._run_output_guardrails(speaker, reply)
+
+                if guardrails_reply is not None:
+                    # if a guardrail has been activated, then the next target has been set and the guardrail reply will be sent
+                    guardrails_activated = True
+                    reply = guardrails_reply
+
             # check for "clear history" phrase in reply and activate clear history function if found
             if (
                 groupchat.enable_clear_history
@@ -1233,7 +1283,11 @@ class GroupChatManager(ConversableAgent):
                 a.previous_cache = None
 
         if termination_reason:
-            iostream.send(TerminationEvent(termination_reason=termination_reason))
+            iostream.send(
+                TerminationEvent(
+                    termination_reason=termination_reason, sender=self, recipient=speaker if speaker else None
+                )
+            )
 
         return True, None
 
@@ -1287,8 +1341,19 @@ class GroupChatManager(ConversableAgent):
             try:
                 # select the next speaker
                 speaker = await groupchat.a_select_speaker(speaker, self)
-                # let the speaker speak
-                reply = await speaker.a_generate_reply(sender=self)
+                if not silent:
+                    iostream.send(GroupChatRunChatEvent(speaker=speaker, silent=silent))
+
+                guardrails_activated = False
+                guardrails_reply = groupchat._run_input_guardrails(speaker, speaker._oai_messages[self])
+
+                if guardrails_reply is not None:
+                    # if a guardrail has been activated, then the next target has been set and the guardrail reply will be sent
+                    guardrails_activated = True
+                    reply = guardrails_reply
+                else:
+                    # let the speaker speak
+                    reply = await speaker.a_generate_reply(sender=self)
             except KeyboardInterrupt:
                 # let the admin agent speak if interrupted
                 if groupchat.admin_name in groupchat.agent_names:
@@ -1308,6 +1373,24 @@ class GroupChatManager(ConversableAgent):
                 termination_reason = "No reply generated"
                 break
 
+            if not guardrails_activated:
+                # if the input guardrails were not activated, and the agent returned a reply
+                guardrails_reply = groupchat._run_output_guardrails(speaker, reply)
+
+                if guardrails_reply is not None:
+                    # if a guardrail has been activated, then the next target has been set and the guardrail reply will be sent
+                    guardrails_activated = True
+                    reply = guardrails_reply
+
+            # check for "clear history" phrase in reply and activate clear history function if found
+            if (
+                groupchat.enable_clear_history
+                and isinstance(reply, dict)
+                and reply["content"]
+                and "CLEAR HISTORY" in reply["content"].upper()
+            ):
+                reply["content"] = self.clear_agents_history(reply, groupchat)
+
             # The speaker sends the message without requesting a reply
             await speaker.a_send(reply, self, request_reply=False, silent=silent)
             message = self.last_message(speaker)
@@ -1317,7 +1400,11 @@ class GroupChatManager(ConversableAgent):
                 a.previous_cache = None
 
         if termination_reason:
-            iostream.send(TerminationEvent(termination_reason=termination_reason))
+            iostream.send(
+                TerminationEvent(
+                    termination_reason=termination_reason, sender=self, recipient=speaker if speaker else None
+                )
+            )
 
         return True, None
 

autogen/agentchat/realtime/experimental/realtime_swarm.py

@@ -359,6 +359,8 @@ class SwarmableRealtimeAgent(SwarmableAgent):
             description=None,
             silent=None,
         )
+        self.input_guardrails = []
+        self.output_guardrails = []
 
     def reset_answer(self) -> None:
         """Reset the answer event."""

autogen/agents/experimental/websurfer/websurfer.py

@@ -12,7 +12,9 @@ from ....tools.experimental import (
     BrowserUseTool,
     Crawl4AITool,
     DuckDuckGoSearchTool,
+    FirecrawlTool,
     PerplexitySearchTool,
+    SearxngSearchTool,
     TavilySearchTool,
 )
 
@@ -28,7 +30,9 @@ class WebSurferAgent(ConversableAgent):
         *,
         llm_config: Optional[Union[LLMConfig, dict[str, Any]]] = None,
         web_tool_llm_config: Optional[Union[LLMConfig, dict[str, Any]]] = None,
-        web_tool: Literal["browser_use", "crawl4ai", "duckduckgo", "perplexity", "tavily"] = "browser_use",
+        web_tool: Literal[
+            "browser_use", "crawl4ai", "duckduckgo", "firecrawl", "perplexity", "tavily", "searxng"
+        ] = "browser_use",
         web_tool_kwargs: Optional[dict[str, Any]] = None,
         **kwargs: Any,
     ) -> None:
@@ -48,12 +52,16 @@ class WebSurferAgent(ConversableAgent):
             self.tool: Tool = BrowserUseTool(llm_config=web_tool_llm_config, **web_tool_kwargs)  # type: ignore[arg-type]
         elif web_tool == "crawl4ai":
             self.tool = Crawl4AITool(llm_config=web_tool_llm_config, **web_tool_kwargs)
+        elif web_tool == "firecrawl":
+            self.tool = FirecrawlTool(llm_config=web_tool_llm_config, **web_tool_kwargs)
         elif web_tool == "perplexity":
             self.tool = PerplexitySearchTool(**web_tool_kwargs)
         elif web_tool == "tavily":
             self.tool = TavilySearchTool(llm_config=web_tool_llm_config, **web_tool_kwargs)
         elif web_tool == "duckduckgo":
             self.tool = DuckDuckGoSearchTool(**web_tool_kwargs)
+        elif web_tool == "searxng":
+            self.tool = SearxngSearchTool(**web_tool_kwargs)
         else:
             raise ValueError(f"Unsupported {web_tool=}.")
 

autogen/code_utils.py

@@ -72,18 +72,20 @@ def content_str(content: Union[str, list[Union[UserMessageTextContentPart, UserM
     if not isinstance(content, list):
         raise TypeError(f"content must be None, str, or list, but got {type(content)}")
 
-    rst = ""
+    rst = []
     for item in content:
         if not isinstance(item, dict):
             raise TypeError("Wrong content format: every element should be dict if the content is a list.")
         assert "type" in item, "Wrong content format. Missing 'type' key in content's dict."
-        if item["type"] == "text":
-            rst += item["text"]
-        elif item["type"] == "image_url":
-            rst += "<image>"
+        if item["type"] in ["text", "input_text"]:
+            rst.append(item["text"])
+        elif item["type"] in ["image_url", "input_image"]:
+            rst.append("<image>")
+        elif item["type"] in ["function", "tool_call", "tool_calls"]:
+            rst.append("<function>" if "name" not in item else f"<function: {item['name']}>")
         else:
             raise ValueError(f"Wrong content format: unknown type {item['type']} within the content")
-    return rst
+    return "\n".join(rst)
 
 
 def infer_lang(code: str) -> str:

autogen/events/agent_events.py

@@ -669,16 +669,22 @@ class TerminationEvent(BaseEvent):
     """When a workflow termination condition is met"""
 
     termination_reason: str
+    sender: str
+    recipient: Optional[str] = None
 
     def __init__(
         self,
         *,
         uuid: Optional[UUID] = None,
+        sender: Union["Agent", str],
+        recipient: Optional[Union["Agent", str]] = None,
         termination_reason: str,
     ):
         super().__init__(
             uuid=uuid,
             termination_reason=termination_reason,
+            sender=sender.name if hasattr(sender, "name") else sender,
+            recipient=recipient.name if hasattr(recipient, "name") else recipient if recipient else None,
         )
 
     def print(self, f: Optional[Callable[..., Any]] = None) -> None:

autogen/events/helpers.py

@@ -13,12 +13,15 @@ logger = logging.getLogger(__name__)
 def deprecated_by(
     new_class: type[BaseModel],
     param_mapping: dict[str, str] = None,
+    default_params: dict[str, any] = None,
 ) -> Callable[[type[BaseModel]], Callable[..., BaseModel]]:
     param_mapping = param_mapping or {}
+    default_params = default_params or {}
 
     def decorator(
         old_class: type[BaseModel],
         param_mapping: dict[str, str] = param_mapping,
+        default_params: dict[str, any] = default_params,
     ) -> Callable[..., BaseModel]:
         @wraps(old_class)
         def wrapper(*args, **kwargs) -> BaseModel:
@@ -28,6 +31,11 @@ def deprecated_by(
             # Translate old parameters to new parameters
             new_kwargs = {param_mapping.get(k, k): v for k, v in kwargs.items()}
 
+            # Add default parameters if not already present
+            for key, value in default_params.items():
+                if key not in new_kwargs:
+                    new_kwargs[key] = value
+
             # Pass the translated parameters to the new class
             return new_class(*args, **new_kwargs)
 

autogen/mcp/helpers.py

@@ -0,0 +1,45 @@
+# Copyright (c) 2023 - 2025, AG2ai, Inc., AG2ai open-source projects maintainers and core contributors
+#
+# SPDX-License-Identifier: Apache-2.0
+import asyncio
+import os
+import signal
+from asyncio.subprocess import PIPE, Process, create_subprocess_exec
+from contextlib import asynccontextmanager
+from typing import AsyncGenerator, Dict, Optional
+
+
+@asynccontextmanager
+async def run_streamable_http_client(
+    *, mcp_server_path: str, env_vars: Optional[Dict[str, str]] = None, startup_wait_secs: float = 5.0
+) -> AsyncGenerator[Process, None]:
+    """
+    Async context manager to run a Python subprocess for streamable-http with custom env vars.
+
+    Args:
+        mcp_server_path: Path to the Python script to run.
+        env_vars: Environment variables to export to the subprocess.
+        startup_wait_secs: Time to wait for the server to start (in seconds).
+    Yields:
+        An asyncio.subprocess.Process object.
+    """
+    env = os.environ.copy()
+    if env_vars:
+        env.update(env_vars)
+
+    process = await create_subprocess_exec(
+        "python", mcp_server_path, "streamable-http", env=env, stdout=PIPE, stderr=PIPE
+    )
+
+    # Optional startup delay to let the server initialize
+    await asyncio.sleep(startup_wait_secs)
+
+    try:
+        yield process
+    finally:
+        if process.returncode is None:
+            process.send_signal(signal.SIGINT)
+            try:
+                await asyncio.wait_for(process.wait(), timeout=5.0)
+            except asyncio.TimeoutError:
+                process.kill()

autogen/mcp/mcp_proxy/mcp_proxy.py

@@ -117,9 +117,8 @@ class MCPProxy:
 
         return q_params, path_params, body, security
 
-    @property
-    def mcp(self) -> "FastMCP":
-        mcp = FastMCP(title=self._title)
+    def get_mcp(self, **settings: Any) -> "FastMCP":
+        mcp = FastMCP(title=self._title, **settings)
 
         for func in self._registered_funcs:
             try:

autogen/messages/agent_messages.py

@@ -647,7 +647,7 @@ class UsingAutoReplyMessage(BaseMessage):
         f(colored("\n>>>>>>>> USING AUTO REPLY...", "red"), flush=True)
 
 
-@deprecated_by(TerminationEvent)
+@deprecated_by(TerminationEvent, default_params={"sender": "system"})
 @wrap_message
 class TerminationMessage(BaseMessage):
     """When a workflow termination condition is met"""

autogen/oai/client.py

@@ -51,6 +51,8 @@ if openai_result.is_successful:
     from openai.types.completion import Completion
     from openai.types.completion_usage import CompletionUsage
 
+    from autogen.oai.openai_responses import OpenAIResponsesClient
+
     if openai.__version__ >= "1.1.0":
         TOOL_ENABLED = True
     ERROR = None
@@ -314,7 +316,7 @@ class ModelClient(Protocol):
     class ModelClientResponseProtocol(Protocol):
         class Choice(Protocol):
             class Message(Protocol):
-                content: Optional[str]
+                content: Optional[str] | Optional[dict[str, Any]]
 
             message: Message
 
@@ -952,6 +954,15 @@ class OpenAIWrapper:
                     raise ImportError("Please install `boto3` to use the Amazon Bedrock API.")
                 client = BedrockClient(response_format=response_format, **openai_config)
                 self._clients.append(client)
+            elif api_type is not None and api_type.startswith("responses"):
+                # OpenAI Responses API (stateful). Reuse the same OpenAI SDK but call the `/responses` endpoint via the new client.
+                @require_optional_import("openai>=1.66.2", "openai")
+                def create_responses_client() -> "OpenAI":
+                    client = OpenAI(**openai_config)
+                    self._clients.append(OpenAIResponsesClient(client, response_format=response_format))
+                    return client
+
+                client = create_responses_client()
             else:
 
                 @require_optional_import("openai>=1.66.2", "openai")
@@ -1442,3 +1453,35 @@ class OpenAIWrapper:
             A list of text, or a list of ChatCompletion objects if function_call/tool_calls are present.
         """
         return response.message_retrieval_function(response)
+
+
+# -----------------------------------------------------------------------------
+# New: Responses API config entry (OpenAI-hosted preview endpoint)
+# -----------------------------------------------------------------------------
+
+
+@register_llm_config
+class OpenAIResponsesLLMConfigEntry(OpenAILLMConfigEntry):
+    """LLMConfig entry for the OpenAI Responses API (stateful, tool-enabled).
+
+    This reuses all the OpenAI fields but changes *api_type* so the wrapper can
+    route traffic to the `client.responses` endpoint instead of
+    `chat.completions`.  It inherits everything else – including reasoning
+    fields – from *OpenAILLMConfigEntry* so users can simply set
+
+    ```python
+    {
+        "api_type": "responses",  # <-- key differentiator
+        "model": "o3",  # reasoning model
+        "reasoning_effort": "medium",  # low / medium / high
+        "stream": True,
+    }
+    ```
+    """
+
+    api_type: Literal["responses"] = "responses"
+    tool_choice: Optional[Literal["none", "auto", "required"]] = "auto"
+    built_in_tools: Optional[list[str]] = None
+
+    def create_client(self) -> "ModelClient":  # pragma: no cover
+        raise NotImplementedError("Handled via OpenAIWrapper._register_default_client")

autogen/oai/gemini.py

@@ -54,7 +54,6 @@ from io import BytesIO
 from typing import Any, Literal, Optional, Type, Union
 
 import requests
-from packaging import version
 from pydantic import BaseModel, Field
 
 from ..import_utils import optional_import_block, require_optional_import
@@ -332,6 +331,12 @@ class GeminiClient:
                 recitation_part = Part(text="Unsuccessful Finish Reason: RECITATION")
                 parts = [recitation_part]
                 error_finish_reason = "content_filter"  # As per available finish_reason in Choice
+            elif not response.candidates[0].content or not response.candidates[0].content.parts:
+                error_part = Part(
+                    text=f"Unsuccessful Finish Reason: ({str(response.candidates[0].finish_reason)}) NO CONTENT RETURNED"
+                )
+                parts = [error_part]
+                error_finish_reason = "content_filter"  # No other option in Choice in chat_completion.py
             else:
                 parts = response.candidates[0].content.parts
         elif isinstance(response, VertexAIGenerationResponse):  # or hasattr(response, "candidates"):
@@ -570,25 +575,19 @@ class GeminiClient:
 
             if part_type == "text":
                 rst.append(
-                    VertexAIContent(parts=parts, role=role)
-                    if self.use_vertexai
-                    else rst.append(Content(parts=parts, role=role))
+                    VertexAIContent(parts=parts, role=role) if self.use_vertexai else Content(parts=parts, role=role)
                 )
-            elif part_type == "tool":
-                # Function responses should be assigned "model" role to keep them separate from function calls
-                role = "function" if version.parse(genai.__version__) < version.parse("1.4.0") else "model"
+            elif part_type == "tool_call":
+                # Function calls should be from the model/assistant
+                role = "model"
                 rst.append(
-                    VertexAIContent(parts=parts, role=role)
-                    if self.use_vertexai
-                    else rst.append(Content(parts=parts, role=role))
+                    VertexAIContent(parts=parts, role=role) if self.use_vertexai else Content(parts=parts, role=role)
                 )
-            elif part_type == "tool_call":
-                # Function calls should be assigned "user" role
-                role = "function" if version.parse(genai.__version__) < version.parse("1.4.0") else "user"
+            elif part_type == "tool":
+                # Function responses should be from the user
+                role = "user"
                 rst.append(
-                    VertexAIContent(parts=parts, role=role)
-                    if self.use_vertexai
-                    else rst.append(Content(parts=parts, role=role))
+                    VertexAIContent(parts=parts, role=role) if self.use_vertexai else Content(parts=parts, role=role)
                 )
             elif part_type == "image":
                 # Image has multiple parts, some can be text and some can be image based
@@ -608,14 +607,14 @@ class GeminiClient:
                     rst.append(
                         VertexAIContent(parts=text_parts, role=role)
                         if self.use_vertexai
-                        else rst.append(Content(parts=text_parts, role=role))
+                        else Content(parts=text_parts, role=role)
                     )
 
                 if len(image_parts) > 0:
                     rst.append(
                         VertexAIContent(parts=image_parts, role=role)
                         if self.use_vertexai
-                        else rst.append(Content(parts=image_parts, role=role))
+                        else Content(parts=image_parts, role=role)
                     )
 
             if len(rst) != 0 and rst[-1] is None:
@@ -908,18 +907,25 @@ def calculate_gemini_cost(use_vertexai: bool, input_tokens: int, output_tokens:
         # https://cloud.google.com/vertex-ai/generative-ai/pricing#vertex-ai-pricing
 
         if (
-            "gemini-2.5-pro-preview-03-25" in model_name
-            or "gemini-2.5-pro-exp-03-25" in model_name
+            model_name == "gemini-2.5-pro"
+            or "gemini-2.5-pro-preview-06-05" in model_name
             or "gemini-2.5-pro-preview-05-06" in model_name
+            or "gemini-2.5-pro-preview-03-25" in model_name
         ):
             if up_to_200k:
                 return total_cost_mil(1.25, 10)
             else:
                 return total_cost_mil(2.5, 15)
 
-        elif "gemini-2.5-flash-preview-04-17" in model_name:
+        elif "gemini-2.5-flash" in model_name:
+            return total_cost_mil(0.3, 2.5)
+
+        elif "gemini-2.5-flash-preview-04-17" in model_name or "gemini-2.5-flash-preview-05-20" in model_name:
             return total_cost_mil(0.15, 0.6)  # NON-THINKING OUTPUT PRICE, $3 FOR THINKING!
 
+        elif "gemini-2.5-flash-lite-preview-06-17" in model_name:
+            return total_cost_mil(0.1, 0.4)
+
         elif "gemini-2.0-flash-lite" in model_name:
             return total_cost_mil(0.075, 0.3)
 
@@ -952,9 +958,10 @@ def calculate_gemini_cost(use_vertexai: bool, input_tokens: int, output_tokens:
         # Non-Vertex AI pricing
 
         if (
-            "gemini-2.5-pro-preview-03-25" in model_name
-            or "gemini-2.5-pro-exp-03-25" in model_name
+            model_name == "gemini-2.5-pro"
+            or "gemini-2.5-pro-preview-06-05" in model_name
             or "gemini-2.5-pro-preview-05-06" in model_name
+            or "gemini-2.5-pro-preview-03-25" in model_name
         ):
             # https://ai.google.dev/gemini-api/docs/pricing#gemini-2.5-pro-preview
             if up_to_200k:
@@ -962,10 +969,18 @@ def calculate_gemini_cost(use_vertexai: bool, input_tokens: int, output_tokens:
             else:
                 return total_cost_mil(2.5, 15)
 
-        elif "gemini-2.5-flash-preview-04-17" in model_name:
+        elif "gemini-2.5-flash" in model_name:
+            # https://ai.google.dev/gemini-api/docs/pricing#gemini-2.5-flash
+            return total_cost_mil(0.3, 2.5)
+
+        elif "gemini-2.5-flash-preview-04-17" in model_name or "gemini-2.5-flash-preview-05-20" in model_name:
             # https://ai.google.dev/gemini-api/docs/pricing#gemini-2.5-flash
             return total_cost_mil(0.15, 0.6)
 
+        elif "gemini-2.5-flash-lite-preview-06-17" in model_name:
+            # https://ai.google.dev/gemini-api/docs/pricing#gemini-2.5-flash-lite
+            return total_cost_mil(0.1, 0.4)
+
         elif "gemini-2.0-flash-lite" in model_name:
             # https://ai.google.dev/gemini-api/docs/pricing#gemini-2.0-flash-lite
             return total_cost_mil(0.075, 0.3)

autogen/oai/gemini_types.py

@@ -91,7 +91,7 @@ class CaseInSensitiveEnum(str, enum.Enum):
                 try:
                     # Creating a enum instance based on the value
                     # We need to use super() to avoid infinite recursion.
-                    unknown_enum_val = super().__new__(cls, value)
+                    unknown_enum_val = super(CaseInSensitiveEnum, cls).__new__(cls, value)
                     unknown_enum_val._name_ = str(value)  # pylint: disable=protected-access
                     unknown_enum_val._value_ = value  # pylint: disable=protected-access
                     return unknown_enum_val