Support xhigh/max/other reasoning efforts + fastmcp 3.x

docs/llm_endpoint_config.md

@@ -110,6 +110,10 @@ agent_model:
     timeout: 60.0
 ```
 
+Reasoning effort is a provider-defined string that is forwarded unchanged.
+Supported values, such as `low`, `medium`, `high`, or `xhigh`, depend on the
+selected model deployment.
+
 **OpenAI-compatible Responses**
 
 Provide the URL directory in `endpoint_url`, model name in `deployment` and do not set `account_name`.

pyproject.toml

@@ -9,7 +9,7 @@ description = "A self-contained agent that thinks inside the box"
 requires-python = ">=3.10"
 dependencies = [
     "mcp[cli]",
-    "fastmcp",
+    "fastmcp>=3.4.7,<4",
     "httpx",
     "pydantic",
     "pydantic[email]",

tests/test_aoai_responses_session.py

@@ -9,6 +9,7 @@
 
 from thinkingbox.common.aoai_responses_session import AOAIResponsesSession
 from thinkingbox.common.chat_types import Text
+from thinkingbox.common.config_types import AOAIResponsesSessionConfig
 
 
 def _mock_client_factory(response_payload):
@@ -48,6 +49,16 @@ async def post(self, url, *args, **kwargs):
 }
 
 
+@pytest.mark.parametrize("reasoning_effort", ["xhigh", "provider-defined"])
+def test_config_accepts_provider_defined_reasoning_effort(reasoning_effort):
+    config = AOAIResponsesSessionConfig(
+        deployment="test",
+        reasoning_effort=reasoning_effort,
+    )
+
+    assert config.reasoning_effort == reasoning_effort
+
+
 @pytest.mark.asyncio
 async def test_response_schema_in_payload(monkeypatch):
     """response_schema should build the full text.format in the request payload."""
@@ -83,6 +94,25 @@ async def test_response_schema_in_payload(monkeypatch):
     }
 
 
+@pytest.mark.asyncio
+async def test_xhigh_reasoning_effort_in_payload(monkeypatch):
+    session = AOAIResponsesSession(
+        deployment="test",
+        endpoint_url="https://test",
+        is_reasoning=True,
+        reasoning_source="none",
+        reasoning_effort="xhigh",
+    )
+    mock_client = _mock_client_factory(SIMPLE_RESPONSE)
+    monkeypatch.setattr(session, "get_client", partial(lambda c: c, mock_client))
+
+    await session.get_completion(
+        conversation=[Text(role="user", content="Hello")],
+    )
+
+    assert mock_client.last_post_kwargs["json"]["reasoning"] == {"effort": "xhigh"}
+
+
 @pytest.mark.asyncio
 async def test_response_schema_requires_explicit_conversation():
     """response_schema without explicit conversation should raise ValueError."""

tests/test_aoai_session.py

@@ -23,6 +23,16 @@ def get_test_aoai_config():
     )
 
 
+@pytest.mark.parametrize("reasoning_effort", ["xhigh", "provider-defined"])
+def test_config_accepts_provider_defined_reasoning_effort(reasoning_effort):
+    config = AOAISessionConfig(
+        deployment="test_deployment",
+        reasoning_effort=reasoning_effort,
+    )
+
+    assert config.reasoning_effort == reasoning_effort
+
+
 SIMPLE_RESPONSE = {
     "choices": [{"message": {"role": "assistant", "content": "Test response"}}]
 }
@@ -148,6 +158,23 @@ async def test_response_schema_in_payload(mock_async_client):
     }
 
 
+@pytest.mark.asyncio
+@patch("thinkingbox.common.llm_session_base.httpx.AsyncClient")
+async def test_xhigh_reasoning_effort_in_payload(mock_async_client):
+    mock_async_client.return_value = get_mock_async_client(SIMPLE_RESPONSE)
+    config = get_test_aoai_config().model_copy(
+        update={"is_reasoning": True, "reasoning_effort": "xhigh"}
+    )
+    session = AOAISession.from_config(config)
+
+    await session._get_completion()
+
+    post_mock = mock_async_client.return_value.__aenter__.return_value.post
+    call_kwargs = post_mock.call_args
+    payload = call_kwargs.kwargs.get("json") or call_kwargs[1]["json"]
+    assert payload["reasoning_effort"] == "xhigh"
+
+
 @pytest.mark.asyncio
 async def test_response_schema_requires_explicit_conversation():
     """response_schema without explicit conversation should raise ValueError."""

thinkingbox/common/aoai_responses_session.py

@@ -4,7 +4,7 @@
 import copy
 import json
 from enum import Enum
-from typing import Any, Literal
+from typing import Any
 
 from thinkingbox.common.chat_types import (
     Message,
@@ -14,7 +14,7 @@
     ToolDef,
     ToolResponse,
 )
-from thinkingbox.common.config_types import AOAIResponsesSessionConfig
+from thinkingbox.common.config_types import AOAIResponsesSessionConfig, ReasoningEffort
 from thinkingbox.common.credential_factory import create_credential
 from thinkingbox.common.llm_session_base import HTTPLLMSessionBase
 from thinkingbox.common.usage_types import Usage
@@ -251,7 +251,7 @@ def _get_conversation_messages(
     async def _responses(
         self,
         messages: list[dict],
-        reasoning_effort_hint: Literal["low", "medium", "high", None],
+        reasoning_effort_hint: ReasoningEffort,
         **kwargs,
     ):
         payload = {

thinkingbox/common/config_types.py

@@ -14,6 +14,8 @@
 
 logger = logging.getLogger(__name__)
 
+ReasoningEffort = str | None
+
 # config.yaml
 
 
@@ -87,7 +89,7 @@ class AOAISessionConfig(HTTPLLMSessionConfig):
     temperature: float = 1.0
     max_completion_tokens: int = 4096
     is_reasoning: bool = False
-    reasoning_effort: Literal["low", "medium", "high", None] = None
+    reasoning_effort: ReasoningEffort = None
     api_version: str = "2024-10-21"
     disabled_params: list[str] = Field(default_factory=list)
     parallel_tool_calls: bool = False
@@ -103,7 +105,7 @@ class AOAIResponsesSessionConfig(HTTPLLMSessionConfig):
     max_completion_tokens: int = 4096
     is_reasoning: bool = False
     reasoning_source: Literal["none", "summary", "content"] = "summary"
-    reasoning_effort: Literal["low", "medium", "high", None] = None
+    reasoning_effort: ReasoningEffort = None
     use_stateful_protocol: bool = False
     api_version: str = "2024-10-21"
     parallel_tool_calls: bool = False