Merge branch 'main' into feature/python-foundry-hosted-agent-vnext

python/packages/core/agent_framework/_tools.py

@@ -906,6 +906,9 @@ def _tools_to_dict(  # pyright: ignore[reportUnusedFunction]
         if isinstance(tool_item, FunctionTool):
             results.append(tool_item.to_json_schema_spec())
             continue
+        if isinstance(tool_item, BaseModel):
+            results.append(tool_item.model_dump(exclude_none=True))
+            continue
         if isinstance(tool_item, SerializationMixin):
             results.append(tool_item.to_dict())
             continue

python/packages/core/agent_framework/_types.py

@@ -1879,6 +1879,12 @@ def _process_update(response: ChatResponse | AgentResponse, update: ChatResponse
             response.finish_reason = update.finish_reason
         if update.model is not None:
             response.model = update.model
+    if (
+        isinstance(response, AgentResponse)
+        and isinstance(update, AgentResponseUpdate)
+        and update.finish_reason is not None
+    ):
+        response.finish_reason = update.finish_reason
     response.continuation_token = update.continuation_token
 
 
@@ -2435,6 +2441,7 @@ class AgentResponse(SerializationMixin, Generic[ResponseModelT]):
         response_id: str | None = None,
         agent_id: str | None = None,
         created_at: CreatedAtT | None = None,
+        finish_reason: FinishReasonLiteral | FinishReason | None = None,
         usage_details: UsageDetails | None = None,
         value: ResponseModelT | None = None,
         response_format: StructuredResponseFormat = None,
@@ -2450,6 +2457,9 @@ class AgentResponse(SerializationMixin, Generic[ResponseModelT]):
             agent_id: The identifier of the agent that produced this response. Useful in multi-agent
                 scenarios to track which agent generated the response.
             created_at: A timestamp for the chat response.
+            finish_reason: The reason the model stopped generating. Common values include
+                ``"stop"`` (natural completion), ``"length"`` (token limit), and
+                ``"tool_calls"`` (the model invoked a tool).
             usage_details: The usage details for the chat response.
             value: The structured output of the agent run response, if applicable.
             response_format: Optional response format for the agent response.
@@ -2476,6 +2486,7 @@ class AgentResponse(SerializationMixin, Generic[ResponseModelT]):
         self.response_id = response_id
         self.agent_id = agent_id
         self.created_at = created_at
+        self.finish_reason = finish_reason
         self.usage_details = usage_details
         self._value: ResponseModelT | None = value
         self._response_format: type[BaseModel] | Mapping[str, Any] | None = response_format
@@ -2688,6 +2699,7 @@ class AgentResponseUpdate(SerializationMixin):
         response_id: str | None = None,
         message_id: str | None = None,
         created_at: CreatedAtT | None = None,
+        finish_reason: FinishReasonLiteral | FinishReason | None = None,
         continuation_token: ContinuationToken | None = None,
         additional_properties: dict[str, Any] | None = None,
         raw_representation: Any | None = None,
@@ -2703,6 +2715,9 @@ class AgentResponseUpdate(SerializationMixin):
             response_id: Optional ID of the response of which this update is a part.
             message_id: Optional ID of the message of which this update is a part.
             created_at: Optional timestamp for the chat response update.
+            finish_reason: The reason the model stopped generating. Common values include
+                ``"stop"`` (natural completion), ``"length"`` (token limit), and
+                ``"tool_calls"`` (the model invoked a tool).
             continuation_token: Optional token for resuming a long-running background operation.
                 When present, indicates the operation is still in progress.
             additional_properties: Optional additional properties associated with the chat response update.
@@ -2729,6 +2744,7 @@ class AgentResponseUpdate(SerializationMixin):
         self.response_id = response_id
         self.message_id = message_id
         self.created_at = created_at
+        self.finish_reason = finish_reason
         self.continuation_token = continuation_token
         self.additional_properties = _restore_compaction_annotation_in_additional_properties(
             additional_properties,
@@ -2761,6 +2777,7 @@ def map_chat_to_agent_update(update: ChatResponseUpdate, agent_name: str | None)
         response_id=update.response_id,
         message_id=update.message_id,
         created_at=update.created_at,
+        finish_reason=update.finish_reason,  # type: ignore[arg-type]
         continuation_token=update.continuation_token,
         additional_properties=update.additional_properties,
         raw_representation=update,

python/packages/core/agent_framework/_workflows/_agent_executor.py

@@ -59,6 +59,62 @@ class AgentExecutorResponse:
     agent_response: AgentResponse
     full_conversation: list[Message]
 
+    def with_text(self, text: str) -> "AgentExecutorResponse":
+        """Create a new AgentExecutorResponse with replaced text, preserving the conversation history.
+
+        Use this in custom executors that transform agent output text (e.g. upper-casing, summarising)
+        when you need downstream AgentExecutors to still have access to the full prior conversation.
+
+        Without this helper, sending a plain ``str`` from a custom executor breaks the context chain:
+        the downstream ``AgentExecutor.from_str`` handler only adds that one string to its cache and
+        loses all prior messages.  By using ``with_text`` the response type stays
+        ``AgentExecutorResponse``, so ``AgentExecutor.from_response`` is invoked instead and the full
+        conversation is preserved.
+
+        Args:
+            text: The replacement assistant message text.
+
+        Returns:
+            A new ``AgentExecutorResponse`` whose ``agent_response`` contains a single assistant
+            message with ``text``, and whose ``full_conversation`` is the prior conversation
+            (everything before the original agent turn) followed by the new assistant message.
+
+        Example:
+            .. code-block:: python
+
+                from agent_framework import AgentExecutorResponse, WorkflowContext, executor
+
+
+                @executor(
+                    id="upper_case_executor",
+                    input=AgentExecutorResponse,
+                    output=AgentExecutorResponse,
+                    workflow_output=str,
+                )
+                async def upper_case(
+                    response: AgentExecutorResponse,
+                    ctx: WorkflowContext[AgentExecutorResponse, str],
+                ) -> None:
+                    upper_text = response.agent_response.text.upper()
+                    await ctx.send_message(response.with_text(upper_text))
+                    await ctx.yield_output(upper_text)
+        """
+        new_message = Message("assistant", [text])
+        new_agent_response = AgentResponse(messages=[new_message])
+
+        # Strip off the original agent turn and replace with the new text.
+        n_agent_messages = len(self.agent_response.messages)
+        prior_messages = (
+            self.full_conversation[:-n_agent_messages] if n_agent_messages else list(self.full_conversation)
+        )
+        new_full_conversation = [*prior_messages, new_message]
+
+        return AgentExecutorResponse(
+            executor_id=self.executor_id,
+            agent_response=new_agent_response,
+            full_conversation=new_full_conversation,
+        )
+
 
 class AgentExecutor(Executor):
     """built-in executor that wraps an agent for handling messages.
@@ -183,7 +239,25 @@ class AgentExecutor(Executor):
         """Accept a raw user prompt string and run the agent.
 
         The new string input will be added to the cache which is used as the conversation context for the agent run.
+
+        Warning:
+            If the upstream executor received an ``AgentExecutorResponse`` but emits a plain
+            ``str``, this handler will be invoked instead of ``from_response``. This resets
+            the conversation context because only the new string is added to the cache and
+            all prior messages from the upstream agent are lost.
+
+            To preserve the full conversation when transforming agent output in a custom
+            executor, use ``AgentExecutorResponse.with_text(...)`` so that the message type
+            stays ``AgentExecutorResponse`` and ``from_response`` is called instead.
         """
+        if not self._cache and ctx.source_executor_ids != ["Workflow"]:
+            logger.warning(
+                "AgentExecutor '%s': from_str handler invoked with an empty cache. "
+                "If you are chaining from an AgentExecutor, the upstream custom executor may be "
+                "emitting a plain str instead of using AgentExecutorResponse.with_text(...), "
+                "which causes the full conversation context to be lost.",
+                self.id,
+            )
         self._cache.extend(normalize_messages_input(text))
         await self._run_agent_and_emit(ctx)
 

python/packages/core/agent_framework/_workflows/_function_executor.py

@@ -268,6 +268,19 @@ def executor(
             forward references. When provided, takes precedence over introspection from the
             ``WorkflowContext`` second generic parameter (W_OutT).
 
+    Warning:
+        When placing a custom ``@executor`` **between** two ``AgentExecutor`` nodes, be
+        careful about the output type.  If the custom executor receives an
+        ``AgentExecutorResponse`` but emits a plain ``str``, the downstream
+        ``AgentExecutor.from_str`` handler is invoked instead of ``from_response``.
+        This resets the conversation context because only the new string is added to
+        the cache and all prior messages from the upstream agent are lost.
+
+        To preserve the full conversation, use
+        ``AgentExecutorResponse.with_text(new_text)`` to create a new response that
+        keeps the prior history, and set ``output=AgentExecutorResponse`` on the
+        decorator.
+
     Returns:
         A FunctionExecutor instance that can be wired into a Workflow.
 

python/packages/core/tests/core/test_tools.py

@@ -16,12 +16,26 @@ from agent_framework._middleware import FunctionInvocationContext
 from agent_framework._tools import (
     _parse_annotation,
     _parse_inputs,
+    _tools_to_dict,
 )
 from agent_framework.observability import OtelAttr
 
 # region FunctionTool and tool decorator tests
 
 
+def test_tools_to_dict_supports_pydantic_tool_models() -> None:
+    """Pydantic-based tool specs are serialized without logging parse warnings."""
+
+    class ProviderTool(BaseModel):
+        kind: str
+        enabled: bool = True
+        note: str | None = None
+
+    result = _tools_to_dict([ProviderTool(kind="google_search")])
+
+    assert result == [{"kind": "google_search", "enabled": True}]
+
+
 def test_tool_decorator():
     """Test the tool decorator."""
 

python/packages/core/tests/core/test_types.py

@@ -40,8 +40,10 @@ from agent_framework._types import (
     _get_data_bytes_as_str,
     _parse_content_list,
     _parse_structured_response_value,
+    _process_update,
     _validate_uri,
     add_usage_details,
+    map_chat_to_agent_update,
     validate_tool_mode,
 )
 from agent_framework.exceptions import AdditionItemMismatch, ContentError
@@ -4179,3 +4181,101 @@ def test_prepend_instructions_custom_role():
 
 
 # endregion
+
+
+# region finish_reason
+
+
+def test_agent_response_init_with_finish_reason() -> None:
+    """Test that AgentResponse correctly initializes and stores finish_reason."""
+    response = AgentResponse(
+        messages=[Message("assistant", [Content.from_text("test")])],
+        finish_reason="stop",
+    )
+    assert response.finish_reason == "stop"
+
+
+def test_agent_response_update_init_with_finish_reason() -> None:
+    """Test that AgentResponseUpdate correctly initializes and stores finish_reason."""
+    update = AgentResponseUpdate(
+        contents=[Content.from_text("test")],
+        role="assistant",
+        finish_reason="stop",
+    )
+    assert update.finish_reason == "stop"
+
+
+def test_map_chat_to_agent_update_forwards_finish_reason() -> None:
+    """Test that mapping a ChatResponseUpdate with finish_reason forwards it."""
+    chat_update = ChatResponseUpdate(
+        contents=[Content.from_text("test")],
+        finish_reason="length",
+    )
+    agent_update = map_chat_to_agent_update(chat_update, agent_name="test_agent")
+
+    assert agent_update.finish_reason == "length"
+    assert agent_update.author_name == "test_agent"
+
+
+def test_process_update_propagates_finish_reason_to_agent_response() -> None:
+    """Test that _process_update correctly updates an AgentResponse from an AgentResponseUpdate."""
+    response = AgentResponse(messages=[Message("assistant", [Content.from_text("test")])])
+    update = AgentResponseUpdate(
+        contents=[Content.from_text("more text")],
+        role="assistant",
+        finish_reason="stop",
+    )
+
+    # Process the update
+    _process_update(response, update)
+
+    assert response.finish_reason == "stop"
+
+
+def test_process_update_does_not_overwrite_with_none() -> None:
+    """Test that _process_update does not overwrite an existing finish_reason with None."""
+    response = AgentResponse(
+        messages=[Message("assistant", [Content.from_text("test")])],
+        finish_reason="length",
+    )
+    update = AgentResponseUpdate(
+        contents=[Content.from_text("more text")],
+        role="assistant",
+        finish_reason=None,
+    )
+
+    # Process the update
+    _process_update(response, update)
+
+    assert response.finish_reason == "length"
+
+
+def test_agent_response_serialization_includes_finish_reason() -> None:
+    """Test that AgentResponse serializes correctly, including finish_reason."""
+    response = AgentResponse(
+        messages=[Message("assistant", [Content.from_text("test")])],
+        response_id="test_123",
+        finish_reason="stop",
+    )
+
+    # Serialize using the framework's API and verify finish_reason is included.
+    data = response.to_dict()
+    assert "finish_reason" in data
+    assert data["finish_reason"] == "stop"
+
+
+def test_agent_response_update_serialization_includes_finish_reason() -> None:
+    """Test that AgentResponseUpdate serializes correctly, including finish_reason."""
+    update = AgentResponseUpdate(
+        contents=[Content.from_text("test")],
+        role="assistant",
+        response_id="test_456",
+        finish_reason="tool_calls",
+    )
+
+    data = update.to_dict()
+    assert "finish_reason" in data
+    assert data["finish_reason"] == "tool_calls"
+
+
+# endregion

python/packages/core/tests/workflow/test_full_conversation.py

@@ -23,6 +23,7 @@ from agent_framework import (
     WorkflowBuilder,
     WorkflowContext,
     WorkflowRunState,
+    executor,
     handler,
 )
 from agent_framework.orchestrations import SequentialBuilder
@@ -478,3 +479,90 @@ async def test_from_response_preserves_service_session_id() -> None:
     assert result.get_outputs() is not None
 
     assert spy_agent._captured_service_session_id == "resp_PREVIOUS_RUN"  # pyright: ignore[reportPrivateUsage]
+
+
+@executor(
+    id="upper_case_executor",
+    input=AgentExecutorResponse,
+    output=AgentExecutorResponse,
+    workflow_output=str,
+)
+async def _upper_case_executor(
+    response: AgentExecutorResponse,
+    ctx: WorkflowContext[AgentExecutorResponse, str],
+) -> None:
+    upper_text = response.agent_response.text.upper()
+    await ctx.send_message(response.with_text(upper_text))
+    await ctx.yield_output(upper_text)
+
+
+async def test_with_text_preserves_full_conversation_through_custom_executor() -> None:
+    """Custom executor using with_text must preserve the full conversation chain."""
+    # Mirrors the reproduction from issue #5246:
+    # agent1 ("User likes sky red") -> agent2 ("User likes sky blue") -> upper_case -> agent3 ("User likes sky green")
+    agent1 = AgentExecutor(
+        _SimpleAgent(id="agent1", name="ContextAgent1", reply_text="User likes sky red"), id="agent1"
+    )
+    agent2 = AgentExecutor(
+        _SimpleAgent(id="agent2", name="ContextAgent2", reply_text="User likes sky blue"), id="agent2"
+    )
+    agent3 = AgentExecutor(
+        _SimpleAgent(id="agent3", name="ContextAgent3", reply_text="User likes sky green"), id="agent3"
+    )
+    capturer = _CaptureFullConversation(id="capture")
+
+    wf = (
+        WorkflowBuilder(start_executor=agent1, output_executors=[capturer])
+        .add_chain([agent1, agent2, _upper_case_executor, agent3, capturer])
+        .build()
+    )
+
+    result = await wf.run("")
+    payload = next(o for o in result.get_outputs() if isinstance(o, dict))
+
+    # The final agent must see the full conversation: user, agent1, UPPER(agent2), agent3
+    assert payload["roles"] == ["user", "assistant", "assistant", "assistant"]
+    assert payload["texts"][1] == "User likes sky red"
+    assert payload["texts"][2] == "USER LIKES SKY BLUE"
+    assert payload["texts"][3] == "User likes sky green"
+
+
+async def test_with_text_does_not_mutate_original() -> None:
+    """with_text returns a new instance; the original must be unmodified."""
+    original = AgentExecutorResponse(
+        executor_id="test_exec",
+        agent_response=AgentResponse(messages=[Message("assistant", ["original reply"])]),
+        full_conversation=[Message("user", ["prompt"]), Message("assistant", ["original reply"])],
+    )
+
+    new = original.with_text("transformed reply")
+
+    assert new is not original
+    assert new.agent_response.text == "transformed reply"
+    assert new.full_conversation[-1].text == "transformed reply"
+    assert new.full_conversation[-1].role == "assistant"
+    # Original unchanged
+    assert original.agent_response.text == "original reply"
+    assert original.full_conversation[-1].text == "original reply"
+
+
+async def test_with_text_strips_multi_message_agent_turn() -> None:
+    """When the agent turn has multiple messages (tool calls), with_text strips all of them."""
+    tool_call = Message("assistant", ["<tool_call>"])
+    tool_result = Message("tool", ["<result>"])
+    final_reply = Message("assistant", ["actual answer"])
+    user_msg = Message("user", ["question"])
+
+    original = AgentExecutorResponse(
+        executor_id="exec",
+        agent_response=AgentResponse(messages=[tool_call, tool_result, final_reply]),
+        full_conversation=[user_msg, tool_call, tool_result, final_reply],
+    )
+
+    new = original.with_text("summarised answer")
+
+    # Only the pre-agent-turn messages should remain, plus the replacement
+    assert len(new.full_conversation) == 2
+    assert new.full_conversation[0].text == "question"
+    assert new.full_conversation[1].text == "summarised answer"
+    assert new.agent_response.text == "summarised answer"