nvidia-nat 1.4.0a20251015__py3-none-any.whl → 1.4.0a20251021__py3-none-any.whl

nat/agent/base.py

@@ -102,11 +102,11 @@ class BaseAgent(ABC):
         AIMessage
             The LLM response
         """
-        output_message = ""
+        output_message = []
         async for event in runnable.astream(inputs, config=config):
-            output_message += event.content
+            output_message.append(event.content)
 
-        return AIMessage(content=output_message)
+        return AIMessage(content="".join(output_message))
 
     async def _call_llm(self, llm: Runnable, inputs: dict[str, Any], config: RunnableConfig | None = None) -> AIMessage:
         """

nat/agent/reasoning_agent/reasoning_agent.py

@@ -157,12 +157,12 @@ async def build_reasoning_function(config: ReasoningFunctionConfig, builder: Bui
             prompt = prompt.to_string()
 
             # Get the reasoning output from the LLM
-            reasoning_output = ""
+            reasoning_output = []
 
             async for chunk in llm.astream(prompt):
-                reasoning_output += chunk.content
+                reasoning_output.append(chunk.content)
 
-            reasoning_output = remove_r1_think_tags(reasoning_output)
+            reasoning_output = remove_r1_think_tags("".join(reasoning_output))
 
             output = await downstream_template.ainvoke(input={
                 "input_text": input_text, "reasoning_output": reasoning_output
@@ -200,12 +200,12 @@ async def build_reasoning_function(config: ReasoningFunctionConfig, builder: Bui
             prompt = prompt.to_string()
 
             # Get the reasoning output from the LLM
-            reasoning_output = ""
+            reasoning_output = []
 
             async for chunk in llm.astream(prompt):
-                reasoning_output += chunk.content
+                reasoning_output.append(chunk.content)
 
-            reasoning_output = remove_r1_think_tags(reasoning_output)
+            reasoning_output = remove_r1_think_tags("".join(reasoning_output))
 
             output = await downstream_template.ainvoke(input={
                 "input_text": input_text, "reasoning_output": reasoning_output

nat/agent/register.py

@@ -19,5 +19,6 @@
 from .prompt_optimizer import register as prompt_optimizer
 from .react_agent import register as react_agent
 from .reasoning_agent import reasoning_agent
+from .responses_api_agent import register as responses_api_agent
 from .rewoo_agent import register as rewoo_agent
 from .tool_calling_agent import register as tool_calling_agent

nat/agent/responses_api_agent/__init__.py

@@ -0,0 +1,14 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

nat/agent/responses_api_agent/register.py

@@ -0,0 +1,126 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+import typing
+
+from pydantic import Field
+
+from nat.agent.base import AGENT_LOG_PREFIX
+from nat.builder.builder import Builder
+from nat.builder.framework_enum import LLMFrameworkEnum
+from nat.builder.function_info import FunctionInfo
+from nat.cli.register_workflow import register_function
+from nat.data_models.component_ref import FunctionRef
+from nat.data_models.component_ref import LLMRef
+from nat.data_models.function import FunctionBaseConfig
+from nat.data_models.openai_mcp import OpenAIMCPSchemaTool
+
+logger = logging.getLogger(__name__)
+
+
+class ResponsesAPIAgentWorkflowConfig(FunctionBaseConfig, name="responses_api_agent"):
+    """
+    Defines an NeMo Agent Toolkit function that uses a Responses API
+    Agent performs reasoning inbetween tool calls, and utilizes the
+    tool names and descriptions to select the optimal tool.
+    """
+
+    llm_name: LLMRef = Field(description="The LLM model to use with the agent.")
+    verbose: bool = Field(default=False, description="Set the verbosity of the agent's logging.")
+    nat_tools: list[FunctionRef] = Field(default_factory=list, description="The list of tools to provide to the agent.")
+    mcp_tools: list[OpenAIMCPSchemaTool] = Field(
+        default_factory=list,
+        description="List of MCP tools to use with the agent. If empty, no MCP tools will be used.")
+    builtin_tools: list[dict[str, typing.Any]] = Field(
+        default_factory=list,
+        description="List of built-in tools to use with the agent. If empty, no built-in tools will be used.")
+
+    max_iterations: int = Field(default=15, description="Number of tool calls before stoping the agent.")
+    description: str = Field(default="Agent Workflow", description="The description of this functions use.")
+    parallel_tool_calls: bool = Field(default=False,
+                                      description="Specify whether to allow parallel tool calls in the agent.")
+    handle_tool_errors: bool = Field(
+        default=True,
+        description="Specify ability to handle tool calling errors. If False, tool errors will raise an exception.")
+
+
+@register_function(config_type=ResponsesAPIAgentWorkflowConfig, framework_wrappers=[LLMFrameworkEnum.LANGCHAIN])
+async def responses_api_agent_workflow(config: ResponsesAPIAgentWorkflowConfig, builder: Builder):
+    from langchain_core.messages.human import HumanMessage
+    from langchain_core.runnables import Runnable
+    from langchain_openai import ChatOpenAI
+
+    from nat.agent.tool_calling_agent.agent import ToolCallAgentGraph
+    from nat.agent.tool_calling_agent.agent import ToolCallAgentGraphState
+
+    llm: ChatOpenAI = await builder.get_llm(config.llm_name, wrapper_type=LLMFrameworkEnum.LANGCHAIN)
+    assert llm.use_responses_api, "Responses API Agent requires an LLM that supports the Responses API."
+
+    # Get tools
+    tools = []
+    nat_tools = await builder.get_tools(tool_names=config.nat_tools, wrapper_type=LLMFrameworkEnum.LANGCHAIN)
+    tools.extend(nat_tools)
+    # MCP tools are optional, if provided they will be used by the agent
+    tools.extend([m.model_dump() for m in config.mcp_tools])
+    # Built-in tools are optional, if provided they will be used by the agent
+    tools.extend(config.builtin_tools)
+
+    # Bind tools to LLM
+    if tools:
+        llm: Runnable = llm.bind_tools(tools=tools, parallel_tool_calls=config.parallel_tool_calls, strict=True)
+
+    if config.verbose:
+        logger.info("%s Using LLM: %s with tools: %s", AGENT_LOG_PREFIX, llm.model_name, tools)
+
+    agent = ToolCallAgentGraph(
+        llm=llm,
+        tools=nat_tools,  # MCP and built-in tools are already bound to the LLM and need not be handled by graph
+        detailed_logs=config.verbose,
+        handle_tool_errors=config.handle_tool_errors)
+
+    graph = await agent.build_graph()
+
+    async def _response_fn(input_message: str) -> str:
+        try:
+            # initialize the starting state with the user query
+            input_message = HumanMessage(content=input_message)
+            state = ToolCallAgentGraphState(messages=[input_message])
+
+            # run the Tool Calling Agent Graph
+            state = await graph.ainvoke(state, config={'recursion_limit': (config.max_iterations + 1) * 2})
+            # setting recursion_limit: 4 allows 1 tool call
+            #   - allows the Tool Calling Agent to perform 1 cycle / call 1 single tool,
+            #   - but stops the agent when it tries to call a tool a second time
+
+            # get and return the output from the state
+            state = ToolCallAgentGraphState(**state)
+            output_message = state.messages[-1]  # pylint: disable=E1136
+            content = output_message.content[-1]['text'] if output_message.content and isinstance(
+                output_message.content[-1], dict) and 'text' in output_message.content[-1] else str(
+                    output_message.content)
+            return content
+        except Exception as ex:
+            logger.exception("%s Tool Calling Agent failed with exception: %s", AGENT_LOG_PREFIX, ex, exc_info=ex)
+            if config.verbose:
+                return str(ex)
+            return "I seem to be having a problem."
+
+    try:
+        yield FunctionInfo.from_fn(_response_fn, description=config.description)
+    except GeneratorExit:
+        logger.exception("%s Workflow exited early!", AGENT_LOG_PREFIX, exc_info=True)
+    finally:
+        logger.debug("%s Cleaning up react_agent workflow.", AGENT_LOG_PREFIX)

nat/agent/tool_calling_agent/agent.py

@@ -233,14 +233,10 @@ def create_tool_calling_agent_prompt(config: "ToolCallAgentWorkflowConfig") -> s
     """
     # the Tool Calling Agent prompt can be customized via config option system_prompt and additional_instructions.
 
-    if config.system_prompt:
-        prompt_str = config.system_prompt
-    else:
-        prompt_str = ""
-
-    if config.additional_instructions:
-        prompt_str += f" {config.additional_instructions}"
-
-    if len(prompt_str) > 0:
-        return prompt_str
+    prompt_strs = []
+    for msg in [config.system_prompt, config.additional_instructions]:
+        if msg is not None:
+            prompt_strs.append(msg)
+    if prompt_strs:
+        return " ".join(prompt_strs)
     return None

nat/builder/context.py

@@ -19,6 +19,7 @@ from collections.abc import Awaitable
 from collections.abc import Callable
 from contextlib import contextmanager
 from contextvars import ContextVar
+from functools import cached_property
 
 from nat.builder.intermediate_step_manager import IntermediateStepManager
 from nat.builder.user_interaction_manager import UserInteractionManager
@@ -167,7 +168,7 @@ class Context:
         """
         return UserInteractionManager(self._context_state)
 
-    @property
+    @cached_property
     def intermediate_step_manager(self) -> IntermediateStepManager:
         """
         Retrieves the intermediate step manager instance from the current context state.

nat/builder/intermediate_step_manager.py

@@ -101,7 +101,10 @@ class IntermediateStepManager:
             open_step = self._outstanding_start_steps.pop(payload.UUID, None)
 
             if (open_step is None):
-                logger.warning("Step id %s not found in outstanding start steps", payload.UUID)
+                logger.warning(
+                    "Step id %s not found in outstanding start steps. "
+                    "This may occur if the step was started in a different context or already completed.",
+                    payload.UUID)
                 return
 
             parent_step_id = open_step.step_parent_id
@@ -157,7 +160,8 @@ class IntermediateStepManager:
             if (open_step is None):
                 logger.warning(
                     "Created a chunk for step %s, but no matching start step was found. "
-                    "Chunks must be created with the same ID as the start step.",
+                    "Chunks must be created with the same ID as the start step. "
+                    "This may occur if the step was started in a different context.",
                     payload.UUID)
                 return
 

nat/data_models/api_server.py

@@ -121,7 +121,15 @@ class Message(BaseModel):
     role: UserMessageContentRoleType
 
 
-class ChatRequestOptionals(BaseModel):
+class ChatRequest(BaseModel):
+    """
+    ChatRequest is a data model that represents a request to the NAT chat API.
+    Fully compatible with OpenAI Chat Completions API specification.
+    """
+
+    # Required fields
+    messages: typing.Annotated[list[Message], conlist(Message, min_length=1)]
+
     # Optional fields (OpenAI Chat Completions API compatible)
     model: str | None = Field(default=None, description="name of the model to use")
     frequency_penalty: float | None = Field(default=0.0,
@@ -145,17 +153,6 @@ class ChatRequestOptionals(BaseModel):
     tool_choice: str | dict[str, typing.Any] | None = Field(default=None, description="Controls which tool is called")
     parallel_tool_calls: bool | None = Field(default=True, description="Whether to enable parallel function calling")
     user: str | None = Field(default=None, description="Unique identifier representing end-user")
-
-
-class ChatRequest(ChatRequestOptionals):
-    """
-    ChatRequest is a data model that represents a request to the NAT chat API.
-    Fully compatible with OpenAI Chat Completions API specification.
-    """
-
-    # Required fields
-    messages: typing.Annotated[list[Message], conlist(Message, min_length=1)]
-
     model_config = ConfigDict(extra="allow",
                               json_schema_extra={
                                   "example": {
@@ -197,39 +194,82 @@ class ChatRequest(ChatRequestOptionals):
                            top_p=top_p)
 
 
-class ChatRequestOrMessage(ChatRequestOptionals):
+class ChatRequestOrMessage(BaseModel):
     """
-    ChatRequestOrMessage is a data model that represents either a conversation or a string input.
+    `ChatRequestOrMessage` is a data model that represents either a conversation or a string input.
     This is useful for functions that can handle either type of input.
 
-    `messages` is compatible with the OpenAI Chat Completions API specification.
-
-    `input_string` is a string input that can be used for functions that do not require a conversation.
-    """
+    - `messages` is compatible with the OpenAI Chat Completions API specification.
+    - `input_message` is a string input that can be used for functions that do not require a conversation.
+
+    Note: When `messages` is provided, extra fields are allowed to enable lossless round-trip
+    conversion with ChatRequest. When `input_message` is provided, no extra fields are permitted.
+    """
+    model_config = ConfigDict(
+        extra="allow",
+        json_schema_extra={
+            "examples": [
+                {
+                    "input_message": "What can you do?"
+                },
+                {
+                    "messages": [{
+                        "role": "user", "content": "What can you do?"
+                    }],
+                    "model": "nvidia/nemotron",
+                    "temperature": 0.7
+                },
+            ],
+            "oneOf": [
+                {
+                    "required": ["input_message"],
+                    "properties": {
+                        "input_message": {
+                            "type": "string"
+                        },
+                    },
+                    "additionalProperties": {
+                        "not": True, "errorMessage": 'remove additional property ${0#}'
+                    },
+                },
+                {
+                    "required": ["messages"],
+                    "properties": {
+                        "messages": {
+                            "type": "array"
+                        },
+                    },
+                    "additionalProperties": True
+                },
+            ]
+        },
+    )
 
     messages: typing.Annotated[list[Message] | None, conlist(Message, min_length=1)] = Field(
-        default=None, description="The conversation messages to process.")
+        default=None, description="A non-empty conversation of messages to process.")
 
-    input_string: str | None = Field(default=None, alias="input_message", description="The input message to process.")
+    input_message: str | None = Field(
+        default=None,
+        description="A single input message to process. Useful for functions that do not require a conversation")
 
     @property
     def is_string(self) -> bool:
-        return self.input_string is not None
+        return self.input_message is not None
 
     @property
     def is_conversation(self) -> bool:
         return self.messages is not None
 
     @model_validator(mode="after")
-    def validate_messages_or_input_string(self):
-        if self.messages is not None and self.input_string is not None:
-            raise ValueError("Either messages or input_message/input_string must be provided, not both")
-        if self.messages is None and self.input_string is None:
-            raise ValueError("Either messages or input_message/input_string must be provided")
-        if self.input_string is not None:
-            extra_fields = self.model_dump(exclude={"input_string"}, exclude_none=True, exclude_unset=True)
+    def validate_model(self):
+        if self.messages is not None and self.input_message is not None:
+            raise ValueError("Either messages or input_message must be provided, not both")
+        if self.messages is None and self.input_message is None:
+            raise ValueError("Either messages or input_message must be provided")
+        if self.input_message is not None:
+            extra_fields = self.model_dump(exclude={"input_message"}, exclude_none=True, exclude_unset=True)
             if len(extra_fields) > 0:
-                raise ValueError("no extra fields are permitted when input_message/input_string is provided")
+                raise ValueError("no extra fields are permitted when input_message is provided")
         return self
 
 
@@ -701,9 +741,9 @@ GlobalTypeConverter.register_converter(_string_to_nat_chat_request)
 
 
 def _chat_request_or_message_to_chat_request(data: ChatRequestOrMessage) -> ChatRequest:
-    if data.input_string is not None:
-        return _string_to_nat_chat_request(data.input_string)
-    return ChatRequest(**data.model_dump(exclude={"input_string"}))
+    if data.input_message is not None:
+        return _string_to_nat_chat_request(data.input_message)
+    return ChatRequest(**data.model_dump(exclude={"input_message"}))
 
 
 GlobalTypeConverter.register_converter(_chat_request_or_message_to_chat_request)
@@ -717,7 +757,17 @@ GlobalTypeConverter.register_converter(_chat_request_to_chat_request_or_message)
 
 
 def _chat_request_or_message_to_string(data: ChatRequestOrMessage) -> str:
-    return data.input_string or ""
+    if data.input_message is not None:
+        return data.input_message
+    # Extract content from last message in conversation
+    if data.messages is None:
+        return ""
+    content = data.messages[-1].content
+    if content is None:
+        return ""
+    if isinstance(content, str):
+        return content
+    return str(content)
 
 
 GlobalTypeConverter.register_converter(_chat_request_or_message_to_string)

nat/data_models/intermediate_step.py

@@ -103,11 +103,19 @@ class ToolSchema(BaseModel):
     function: ToolDetails = Field(..., description="The function details.")
 
 
+class ServerToolUseSchema(BaseModel):
+    name: str
+    arguments: str | dict[str, typing.Any] | typing.Any
+    output: typing.Any
+
+    model_config = ConfigDict(extra="ignore")
+
+
 class TraceMetadata(BaseModel):
     chat_responses: typing.Any | None = None
     chat_inputs: typing.Any | None = None
     tool_inputs: typing.Any | None = None
-    tool_outputs: typing.Any | None = None
+    tool_outputs: list[ServerToolUseSchema] | typing.Any | None = None
     tool_info: typing.Any | None = None
     span_inputs: typing.Any | None = None
     span_outputs: typing.Any | None = None

nat/data_models/llm.py

@@ -14,14 +14,28 @@
 # limitations under the License.
 
 import typing
+from enum import Enum
+
+from pydantic import Field
 
 from .common import BaseModelRegistryTag
 from .common import TypedBaseModel
 
 
+class APITypeEnum(str, Enum):
+    CHAT_COMPLETION = "chat_completion"
+    RESPONSES = "responses"
+
+
 class LLMBaseConfig(TypedBaseModel, BaseModelRegistryTag):
     """Base configuration for LLM providers."""
-    pass
+
+    api_type: APITypeEnum = Field(default=APITypeEnum.CHAT_COMPLETION,
+                                  description="The type of API to use for the LLM provider.",
+                                  json_schema_extra={
+                                      "enum": [e.value for e in APITypeEnum],
+                                      "examples": [e.value for e in APITypeEnum],
+                                  })
 
 
 LLMBaseConfigT = typing.TypeVar("LLMBaseConfigT", bound=LLMBaseConfig)

nat/data_models/openai_mcp.py

@@ -0,0 +1,46 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from enum import Enum
+
+from pydantic import BaseModel
+from pydantic import ConfigDict
+from pydantic import Field
+
+
+class MCPApprovalRequiredEnum(str, Enum):
+    """
+    Enum to specify if approval is required for tool usage in the OpenAI MCP schema.
+    """
+    NEVER = "never"
+    ALWAYS = "always"
+    AUTO = "auto"
+
+
+class OpenAIMCPSchemaTool(BaseModel):
+    """
+    Represents a tool in the OpenAI MCP schema.
+    """
+    type: str = "mcp"
+    server_label: str = Field(description="Label for the server where the tool is hosted.")
+    server_url: str = Field(description="URL of the server hosting the tool.")
+    allowed_tools: list[str] | None = Field(default=None,
+                                            description="List of allowed tool names that can be used by the agent.")
+    require_approval: MCPApprovalRequiredEnum = Field(default=MCPApprovalRequiredEnum.NEVER,
+                                                      description="Specifies if approval is required for tool usage.")
+    headers: dict[str, str] | None = Field(default=None,
+                                           description="Optional headers to include in requests to the tool server.")
+
+    model_config = ConfigDict(use_enum_values=True)

nat/data_models/optimizable.py

@@ -23,6 +23,7 @@ from pydantic import BaseModel
 from pydantic import ConfigDict
 from pydantic import Field
 from pydantic import model_validator
+from pydantic_core import PydanticUndefined
 
 T = TypeVar("T", int, float, bool, str)
 
@@ -66,7 +67,7 @@ class SearchSpace(BaseModel, Generic[T]):
 
 
 def OptimizableField(
-    default: Any,
+    default: Any = PydanticUndefined,
     *,
     space: SearchSpace | None = None,
     merge_conflict: str = "overwrite",

nat/data_models/thinking_mixin.py

@@ -51,7 +51,7 @@ class ThinkingMixin(
         Returns the system prompt to use for thinking.
         For NVIDIA Nemotron, returns "/think" if enabled, else "/no_think".
         For Llama Nemotron v1.5, returns "/think" if enabled, else "/no_think".
-        For Llama Nemotron v1.0, returns "detailed thinking on" if enabled, else "detailed thinking off".
+        For Llama Nemotron v1.0 or v1.1, returns "detailed thinking on" if enabled, else "detailed thinking off".
         If thinking is not supported on the model, returns None.
 
         Returns:
@@ -72,7 +72,7 @@ class ThinkingMixin(
                 return "/think" if self.thinking else "/no_think"
 
             if model.startswith("nvidia/llama"):
-                if "v1-0" in model or "v1-1" in model:
+                if "v1-0" in model or "v1-1" in model or model.endswith("v1"):
                     return f"detailed thinking {'on' if self.thinking else 'off'}"
 
                 if "v1-5" in model:

nat/eval/evaluate.py

@@ -104,6 +104,8 @@ class EvaluationRun:
                 usage_stats_per_llm[llm_name].prompt_tokens += step.token_usage.prompt_tokens
                 usage_stats_per_llm[llm_name].completion_tokens += step.token_usage.completion_tokens
                 usage_stats_per_llm[llm_name].total_tokens += step.token_usage.total_tokens
+                usage_stats_per_llm[llm_name].reasoning_tokens += step.token_usage.reasoning_tokens
+                usage_stats_per_llm[llm_name].cached_tokens += step.token_usage.cached_tokens
                 total_tokens += step.token_usage.total_tokens
 
         # find min and max event timestamps

nat/eval/usage_stats.py

@@ -21,6 +21,8 @@ from pydantic import BaseModel
 class UsageStatsLLM(BaseModel):
     prompt_tokens: int = 0
     completion_tokens: int = 0
+    cached_tokens: int = 0
+    reasoning_tokens: int = 0
     total_tokens: int = 0
 
 

nat/front_ends/fastapi/fastapi_front_end_plugin_worker.py

@@ -1184,6 +1184,7 @@ class FastApiFrontEndPluginWorker(FastApiFrontEndPluginWorkerBase):
                             "server": client.server_name,
                             "transport": config.server.transport,
                             "session_healthy": session_healthy,
+                            "protected": True if config.server.auth_provider is not None else False,
                             "tools": tools_info,
                             "total_tools": len(configured_short_names),
                             "available_tools": available_count
@@ -1196,6 +1197,7 @@ class FastApiFrontEndPluginWorker(FastApiFrontEndPluginWorkerBase):
                             "server": "unknown",
                             "transport": config.server.transport if config.server else "unknown",
                             "session_healthy": False,
+                            "protected": False,
                             "error": str(e),
                             "tools": [],
                             "total_tools": 0,
@@ -1226,6 +1228,7 @@ class FastApiFrontEndPluginWorker(FastApiFrontEndPluginWorkerBase):
                                     "server": "streamable-http:http://localhost:9901/mcp",
                                     "transport": "streamable-http",
                                     "session_healthy": True,
+                                    "protected": False,
                                     "tools": [{
                                         "name": "tool_a",
                                         "description": "Tool A description",