agent-framework-core 1.2.0__tar.gz → 1.2.2__tar.gz

{agent_framework_core-1.2.0 → agent_framework_core-1.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agent-framework-core
-Version: 1.2.0
+Version: 1.2.2
 Summary: Microsoft Agent Framework for building AI Agents with Python. This is the core package that has all the core abstractions and implementations.
 Author-email: Microsoft <af-support@microsoft.com>
 Requires-Python: >=3.10

{agent_framework_core-1.2.0 → agent_framework_core-1.2.2}/agent_framework/__init__.py

@@ -247,6 +247,7 @@ from ._workflows._workflow_executor import (
     WorkflowExecutor,
 )
 from .exceptions import (
+    AgentFrameworkException,
     MiddlewareException,
     UserInputRequiredException,
     WorkflowCheckpointException,
@@ -280,6 +281,7 @@ __all__ = [
     "AgentExecutor",
     "AgentExecutorRequest",
     "AgentExecutorResponse",
+    "AgentFrameworkException",
     "AgentMiddleware",
     "AgentMiddlewareLayer",
     "AgentMiddlewareTypes",

{agent_framework_core-1.2.0 → agent_framework_core-1.2.2}/agent_framework/_telemetry.py

@@ -79,9 +79,11 @@ def _detect_hosted_environment() -> None:
     except (ModuleNotFoundError, ValueError):
         return
     with contextlib.suppress(ImportError, AttributeError):
-        from azure.ai.agentserver.core import AgentConfig  # pyright: ignore[reportMissingImports]
+        from azure.ai.agentserver.core import (  # pyright: ignore[reportMissingImports]
+            AgentConfig,  # pyright: ignore[reportUnknownVariableType]
+        )
 
-        if AgentConfig.from_env().is_hosted:
+        if AgentConfig.from_env().is_hosted:  # pyright: ignore[reportUnknownMemberType]
             _add_user_agent_prefix(_HOSTED_USER_AGENT_PREFIX)
             _hosted_env_detected = True
 

{agent_framework_core-1.2.0 → agent_framework_core-1.2.2}/agent_framework/_types.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import base64
+import contextlib
 import json
 import logging
 import re
@@ -2890,6 +2891,7 @@ class ResponseStream(AsyncIterable[UpdateT], Generic[UpdateT, FinalT]):
         self._inner_stream_source: ResponseStream[Any, Any] | Awaitable[ResponseStream[Any, Any]] | None = None
         self._wrap_inner: bool = False
         self._map_update: Callable[[Any], UpdateT | Awaitable[UpdateT]] | None = None
+        self._pull_context_manager_factories: list[Callable[[], contextlib.AbstractContextManager[Any]]] = []
 
     def map(
         self,
@@ -3008,11 +3010,18 @@ class ResponseStream(AsyncIterable[UpdateT], Generic[UpdateT, FinalT]):
         return self
 
     async def __anext__(self) -> UpdateT:
-        if self._iterator is None:
-            stream = await self._get_stream()
-            self._iterator = stream.__aiter__()
         try:
-            update: UpdateT = await self._iterator.__anext__()
+            with contextlib.ExitStack() as stack:
+                for factory in self._pull_context_manager_factories:
+                    stack.enter_context(factory())
+                # Resolve the underlying stream inside the pull contexts so that any
+                # spans/contexts created during stream resolution (e.g. inner chat
+                # completion spans created on the first pull of a wrapped agent stream)
+                # inherit the active context (e.g. an outer agent invoke span).
+                if self._iterator is None:
+                    stream = await self._get_stream()
+                    self._iterator = stream.__aiter__()
+                update: UpdateT = await self._iterator.__anext__()
         except StopAsyncIteration:
             self._consumed = True
             await self._run_cleanup_hooks()
@@ -3038,9 +3047,25 @@ class ResponseStream(AsyncIterable[UpdateT], Generic[UpdateT, FinalT]):
                 update = hooked
         return update
 
+    async def _resolve_stream_with_pull_contexts(self) -> AsyncIterable[UpdateT]:
+        """Resolve the underlying stream while activating any registered pull context managers.
+
+        Used by ``__await__`` and ``get_final_response`` so that any spans/contexts created
+        during stream resolution (e.g. when the source is an Awaitable that internally
+        creates child telemetry spans) inherit the same active context as iterator pulls.
+        ``__anext__`` resolves the stream inside its own ExitStack and so calls ``_get_stream``
+        directly.
+        """
+        if self._stream is not None:
+            return await self._get_stream()
+        with contextlib.ExitStack() as stack:
+            for factory in self._pull_context_manager_factories:
+                stack.enter_context(factory())
+            return await self._get_stream()
+
     def __await__(self) -> Any:
         async def _wrap() -> ResponseStream[UpdateT, FinalT]:
-            await self._get_stream()
+            await self._resolve_stream_with_pull_contexts()
             return self
 
         return _wrap().__await__()
@@ -3064,10 +3089,12 @@ class ResponseStream(AsyncIterable[UpdateT], Generic[UpdateT, FinalT]):
         """
         if self._wrap_inner:
             if self._inner_stream is None:
-                # Use _get_stream() to resolve the awaitable - this properly handles
+                # Use _resolve_stream_with_pull_contexts() so that any spans/contexts
+                # created while resolving the awaitable (e.g. inner telemetry spans)
+                # inherit the same active context as iterator pulls. This also handles
                 # the case where _stream_source and _inner_stream_source are the same
                 # coroutine (e.g., from from_awaitable), avoiding double-await errors.
-                await self._get_stream()
+                await self._resolve_stream_with_pull_contexts()
             if self._inner_stream is None:
                 raise RuntimeError("Inner stream not available")
             if not self._finalized and not self._consumed:
@@ -3177,6 +3204,25 @@ class ResponseStream(AsyncIterable[UpdateT], Generic[UpdateT, FinalT]):
         self._cleanup_hooks.append(hook)
         return self
 
+    def with_pull_context_manager(
+        self,
+        cm_factory: Callable[[], contextlib.AbstractContextManager[Any]],
+    ) -> ResponseStream[UpdateT, FinalT]:
+        """Register a context manager factory invoked around each underlying iterator pull.
+
+        The factory is called once per ``__anext__`` and the returned context manager wraps
+        the await of the underlying iterator. This is useful for state that needs to be
+        active while the inner async work runs - for example, attaching an OpenTelemetry
+        span to the current context so child spans created by inner code (HTTP clients,
+        tool execution) are correctly parented.
+
+        Because the context manager is entered and exited within the same ``__anext__``
+        invocation, attach/detach style operations remain symmetric in the same async
+        context regardless of where the stream is iterated.
+        """
+        self._pull_context_manager_factories.append(cm_factory)
+        return self
+
     async def _run_cleanup_hooks(self) -> None:
         if self._cleanup_run:
             return

{agent_framework_core-1.2.0 → agent_framework_core-1.2.2}/agent_framework/_workflows/_agent.py

@@ -437,6 +437,13 @@ class WorkflowAgent(BaseAgent):
                     yield event
 
         elif checkpoint_id is not None:
+            # Restore the prior workflow state from the checkpoint. Shared
+            # state (e.g. accumulated conversation history maintained by the
+            # workflow's executors) survives across turns because Workflow.run
+            # no longer wipes state per call. Callers who want to deliver a
+            # new user message after restore should make a second
+            # `workflow.run(message=...)` call - they are NOT mutually
+            # exclusive on the same instance, but each must be its own call.
             if streaming:
                 async for event in self.workflow.run(
                     stream=True,
@@ -528,6 +535,7 @@ class WorkflowAgent(BaseAgent):
                 raw_representations.append(output_event)
             else:
                 data = output_event.data
+
                 if isinstance(data, AgentResponseUpdate):
                     # We cannot support AgentResponseUpdate in non-streaming mode. This is because the message
                     # sequence cannot be guaranteed when there are streaming updates in between non-streaming
@@ -628,16 +636,23 @@ class WorkflowAgent(BaseAgent):
             A list of AgentResponseUpdate objects. Empty list if the event is not relevant.
         """
         if event.type == "output":
-            # Convert workflow output to agent response updates.
-            # Handle different data types appropriately.
             data = event.data
             executor_id = event.executor_id
 
             if isinstance(data, AgentResponseUpdate):
-                # Pass through AgentResponseUpdate directly (streaming from AgentExecutor)
-                if not data.author_name:
-                    data.author_name = executor_id
-                return [data]
+                # Construct a fresh AgentResponseUpdate so we don't mutate a payload
+                # that AgentExecutor still holds a reference to in its `updates` list.
+                return [
+                    AgentResponseUpdate(
+                        contents=list(data.contents),
+                        role=data.role,
+                        author_name=data.author_name or executor_id,
+                        response_id=data.response_id,
+                        message_id=data.message_id,
+                        created_at=data.created_at,
+                        raw_representation=data.raw_representation,
+                    )
+                ]
             if isinstance(data, AgentResponse):
                 # Convert each message in AgentResponse to an AgentResponseUpdate
                 updates: list[AgentResponseUpdate] = []

{agent_framework_core-1.2.0 → agent_framework_core-1.2.2}/agent_framework/_workflows/_agent_executor.py

@@ -156,8 +156,9 @@ class AgentExecutor(Executor):
                    the agent run.
                 - "custom": use the provided context_filter function to determine which messages to include
                    as context for the agent run.
-            context_filter: An optional function for filtering conversation context when context_mode is set
-                to "custom".
+            context_filter: A function that takes the full conversation (list of Messages) as input and returns
+                a filtered list of Messages to be used as context for the agent run. This is required
+                if context_mode is set to "custom".
         """
         # Prefer provided id; else use agent.name if present; else generate deterministic prefix
         exec_id = id or resolve_agent_id(agent)

{agent_framework_core-1.2.0 → agent_framework_core-1.2.2}/agent_framework/_workflows/_runner.py

@@ -278,7 +278,12 @@ class Runner:
                     "Please rebuild the original workflow before resuming."
                 )
 
-            # Restore state
+            # Restore state. Clear first so import_state (which merges) does
+            # not leak stale keys from a prior run on this Workflow instance.
+            # This matters more now that Workflow.run() no longer wipes state
+            # per call - the only reset point for shared state on a reused
+            # instance is at restore time.
+            self._state.clear()
             self._state.import_state(checkpoint.state)
             # Restore executor states using the restored state
             await self._restore_executor_states()

{agent_framework_core-1.2.0 → agent_framework_core-1.2.2}/agent_framework/_workflows/_workflow.py

@@ -299,7 +299,7 @@ class Workflow(DictConvertible):
     async def _run_workflow_with_tracing(
         self,
         initial_executor_fn: Callable[[], Awaitable[None]] | None = None,
-        reset_context: bool = True,
+        is_continuation: bool = False,
         streaming: bool = False,
         function_invocation_kwargs: Mapping[str, Mapping[str, Any]] | Mapping[str, Any] | None = None,
         client_kwargs: Mapping[str, Mapping[str, Any]] | Mapping[str, Any] | None = None,
@@ -310,13 +310,19 @@ class Workflow(DictConvertible):
         of external callers to maintain context across different workflow runs.
 
         Args:
-            initial_executor_fn: Optional function to execute initial executor
-            reset_context: Whether to reset the context for a new run
-            streaming: Whether to enable streaming mode for agents
+            initial_executor_fn: Optional function to execute initial executor.
+            is_continuation: True when this run is a continuation of prior
+                work (a checkpoint restore or a responses-only replay) rather
+                than a fresh new turn delivered via the start executor with
+                ``message=...``. Continuations preserve per-run accounting
+                (iteration counter and run kwargs) from the prior turn;
+                fresh-message runs reset them. Shared workflow state is
+                preserved in both cases.
+            streaming: Whether to enable streaming mode for agents.
             function_invocation_kwargs: Optional kwargs to store in State for function
-                invocations in subagents
+                invocations in subagents.
             client_kwargs: Optional kwargs to store in State for chat client
-                invocations in subagents
+                invocations in subagents.
 
         Yields:
             WorkflowEvent: The events generated during the workflow execution.
@@ -345,16 +351,26 @@ class Workflow(DictConvertible):
                     in_progress = WorkflowEvent.status(WorkflowRunState.IN_PROGRESS)
                 yield in_progress  # noqa: RUF070
 
-                # Reset context for a new run if supported
-                if reset_context:
+                # Per-run reset for fresh-message runs only. We deliberately
+                # do NOT clear shared workflow state (`_state.clear()`) or the
+                # runner context's in-flight messages (`reset_for_new_run()`)
+                # here - state and pending work persist across `run()` calls
+                # so that a `WorkflowAgent` can deliver multi-turn input on
+                # the same instance and have prior turns' context survive.
+                # Iteration counting and per-run kwargs ARE per-run though,
+                # so they're reset here.
+                if not is_continuation:
                     self._runner.reset_iteration_count()
-                    self._runner.context.reset_for_new_run()
-                    self._state.clear()
 
                 # Store run kwargs in State so executors can access them.
-                # Only overwrite when new kwargs are explicitly provided or state was
-                # just cleared (fresh run). On continuation (reset_context=False) with
-                # no new kwargs, preserve the kwargs from the original run.
+                # Per-run kwargs semantics:
+                # - On a fresh message run, prior kwargs go away (set to {}
+                #   by default, or to the new kwargs if provided). This
+                #   prevents stale kwargs from a prior turn leaking into the
+                #   current turn.
+                # - On a continuation (checkpoint restore or responses), the
+                #   prior run's kwargs are preserved unless the caller
+                #   explicitly provides new kwargs.
                 if function_invocation_kwargs is not None or client_kwargs is not None:
                     combined_kwargs: dict[str, Any] = {}
                     if function_invocation_kwargs is not None:
@@ -366,11 +382,12 @@ class Workflow(DictConvertible):
                             client_kwargs, "client_kwargs"
                         )
                     self._state.set(WORKFLOW_RUN_KWARGS_KEY, combined_kwargs)
-                elif reset_context:
+                elif not is_continuation:
                     self._state.set(WORKFLOW_RUN_KWARGS_KEY, {})
                 self._state.commit()  # Commit immediately so kwargs are available
 
-                # Set streaming mode after reset
+                # Set streaming mode (always set explicitly per run since
+                # reset_for_new_run() no longer runs to clear it).
                 self._runner_context.set_streaming(streaming)
 
                 # Execute initial setup if provided
@@ -585,13 +602,31 @@ class Workflow(DictConvertible):
         if checkpoint_storage is not None:
             self._runner.context.set_runtime_checkpoint_storage(checkpoint_storage)
 
-        initial_executor_fn, reset_context = self._resolve_execution_mode(
-            message, responses, checkpoint_id, checkpoint_storage
-        )
+        # Async validation: a fresh-message run is only allowed when the
+        # runner context has fully drained from any prior run. If it still
+        # has in-flight executor messages, the prior run didn't complete -
+        # the caller must either resume from a checkpoint or wait for the
+        # prior run to drain. (Pending request_info events are intentionally
+        # NOT blocked here: a follow-up run with message=... is the normal
+        # way to deliver a response to those pending requests, e.g. via
+        # WorkflowAgent._process_pending_requests.)
+        # NOTE: _validate_run_params already enforces that ``message`` is
+        # mutually exclusive with both ``checkpoint_id`` and ``responses``,
+        # so we don't need to re-check those here.
+        if message is not None and await self._runner.context.has_messages():
+            raise RuntimeError(
+                "Cannot start a new run with 'message' while in-flight executor "
+                "messages remain from a prior run. Resume from a checkpoint "
+                "(checkpoint_id=...) or wait for the prior run to complete. "
+                "Workflows that need to recover from a mid-run failure must use "
+                "checkpointing; there is no in-process recovery path."
+            )
+
+        initial_executor_fn = self._resolve_execution_mode(message, responses, checkpoint_id, checkpoint_storage)
 
         async for event in self._run_workflow_with_tracing(
             initial_executor_fn=initial_executor_fn,
-            reset_context=reset_context,
+            is_continuation=(message is None),
             streaming=streaming,
             function_invocation_kwargs=function_invocation_kwargs,
             client_kwargs=client_kwargs,
@@ -674,12 +709,8 @@ class Workflow(DictConvertible):
         responses: Mapping[str, Any] | None,
         checkpoint_id: str | None,
         checkpoint_storage: CheckpointStorage | None,
-    ) -> tuple[Callable[[], Awaitable[None]], bool]:
-        """Determine the initial executor function and reset_context flag based on parameters.
-
-        Returns:
-            A tuple of (initial_executor_fn, reset_context).
-        """
+    ) -> Callable[[], Awaitable[None]]:
+        """Determine the initial executor function based on parameters."""
         if responses is not None:
             if checkpoint_id is not None:
                 # Combined: restore checkpoint then send responses
@@ -689,13 +720,9 @@ class Workflow(DictConvertible):
             else:
                 # Send responses only (requires pending requests in workflow state)
                 initial_executor_fn = functools.partial(self._send_responses_internal, responses)
-            return initial_executor_fn, False
+            return initial_executor_fn
         # Regular run or checkpoint restoration
-        initial_executor_fn = functools.partial(
-            self._execute_with_message_or_checkpoint, message, checkpoint_id, checkpoint_storage
-        )
-        reset_context = message is not None and checkpoint_id is None
-        return initial_executor_fn, reset_context
+        return functools.partial(self._execute_with_message_or_checkpoint, message, checkpoint_id, checkpoint_storage)
 
     async def _restore_and_send_responses(
         self,

{agent_framework_core-1.2.0 → agent_framework_core-1.2.2}/agent_framework/_workflows/_workflow_executor.py

@@ -361,7 +361,7 @@ class WorkflowExecutor(Executor):
         return any(is_instance_of(message.data, input_type) for input_type in self.workflow.input_types)
 
     @handler
-    async def process_workflow(self, input_data: object, ctx: WorkflowContext[Any]) -> None:
+    async def process_workflow(self, input_data: object, ctx: WorkflowContext[Any, Any]) -> None:
         """Execute the sub-workflow with raw input data.
 
         This handler starts a new sub-workflow execution. When the sub-workflow
@@ -428,7 +428,7 @@ class WorkflowExecutor(Executor):
     async def handle_message_wrapped_request_response(
         self,
         response: SubWorkflowResponseMessage,
-        ctx: WorkflowContext[Any],
+        ctx: WorkflowContext[Any, Any],
     ) -> None:
         """Handle response from parent for a forwarded request.
 

{agent_framework_core-1.2.0 → agent_framework_core-1.2.2}/agent_framework/exceptions.py

@@ -34,7 +34,8 @@ class AgentFrameworkException(Exception):
             logger.log(log_level, message, exc_info=inner_exception)
         if inner_exception:
             super().__init__(message, inner_exception, *args)  # type: ignore
-        super().__init__(message, *args)  # type: ignore
+        else:
+            super().__init__(message, *args)  # type: ignore
 
 
 # region Agent Exceptions

{agent_framework_core-1.2.0 → agent_framework_core-1.2.2}/agent_framework/foundry/__init__.py

@@ -4,6 +4,7 @@
 
 This module lazily re-exports objects from:
 - ``agent-framework-anthropic``
+- ``agent-framework-azure-contentunderstanding``
 - ``agent-framework-foundry``
 - ``agent-framework-foundry-local``
 """
@@ -12,7 +13,15 @@ import importlib
 from typing import Any
 
 _IMPORTS: dict[str, tuple[str, str]] = {
+    "AnalysisSection": ("agent_framework_azure_contentunderstanding", "agent-framework-azure-contentunderstanding"),
     "AnthropicFoundryClient": ("agent_framework_anthropic", "agent-framework-anthropic"),
+    "ContentUnderstandingContextProvider": (
+        "agent_framework_azure_contentunderstanding",
+        "agent-framework-azure-contentunderstanding",
+    ),
+    "DocumentStatus": ("agent_framework_azure_contentunderstanding", "agent-framework-azure-contentunderstanding"),
+    "FileSearchBackend": ("agent_framework_azure_contentunderstanding", "agent-framework-azure-contentunderstanding"),
+    "FileSearchConfig": ("agent_framework_azure_contentunderstanding", "agent-framework-azure-contentunderstanding"),
     "FoundryAgent": ("agent_framework_foundry", "agent-framework-foundry"),
     "FoundryAgentOptions": ("agent_framework_foundry", "agent-framework-foundry"),
     "FoundryChatClient": ("agent_framework_foundry", "agent-framework-foundry"),

{agent_framework_core-1.2.0 → agent_framework_core-1.2.2}/agent_framework/foundry/__init__.pyi

@@ -4,6 +4,13 @@
 # Install the relevant packages for full type support.
 
 from agent_framework_anthropic import AnthropicFoundryClient, RawAnthropicFoundryClient
+from agent_framework_azure_contentunderstanding import (  # pyright: ignore[reportMissingImports]
+    AnalysisSection,  # pyright: ignore[reportUnknownVariableType]
+    ContentUnderstandingContextProvider,  # pyright: ignore[reportUnknownVariableType]
+    DocumentStatus,  # pyright: ignore[reportUnknownVariableType]
+    FileSearchBackend,  # pyright: ignore[reportUnknownVariableType]
+    FileSearchConfig,  # pyright: ignore[reportUnknownVariableType]
+)
 from agent_framework_foundry import (
     FoundryAgent,
     FoundryChatClient,
@@ -31,7 +38,12 @@ from agent_framework_foundry_local import (
 )
 
 __all__ = [
+    "AnalysisSection",
     "AnthropicFoundryClient",
+    "ContentUnderstandingContextProvider",
+    "DocumentStatus",
+    "FileSearchBackend",
+    "FileSearchConfig",
     "FoundryAgent",
     "FoundryChatClient",
     "FoundryChatOptions",

{agent_framework_core-1.2.0 → agent_framework_core-1.2.2}/agent_framework/observability.py

@@ -26,6 +26,7 @@ from time import perf_counter, time_ns
 from typing import TYPE_CHECKING, Any, ClassVar, Final, Generic, Literal, TypedDict, cast, overload
 
 from dotenv import load_dotenv
+from opentelemetry import context as otel_context
 from opentelemetry import metrics, trace
 
 from . import __version__ as version_info
@@ -1277,27 +1278,8 @@ class ChatTelemetryLayer(Generic[OptionsCoT]):
         )
 
         if stream:
-            result_stream = cast(
-                ResponseStream[ChatResponseUpdate, ChatResponse[Any]],
-                super_get_response(
-                    messages=messages,
-                    stream=True,
-                    options=opts,
-                    compaction_strategy=compaction_strategy,
-                    tokenizer=tokenizer,
-                    function_invocation_kwargs=function_invocation_kwargs,
-                    client_kwargs=merged_client_kwargs,
-                ),
-            )
+            span = _start_streaming_span(attributes, OtelAttr.REQUEST_MODEL)
 
-            # Create span directly without trace.use_span() context attachment.
-            # Streaming spans are closed asynchronously in cleanup hooks, which run
-            # in a different async context than creation — using use_span() would
-            # cause "Failed to detach context" errors from OpenTelemetry.
-            operation = attributes.get(OtelAttr.OPERATION, "operation")
-            span_name = attributes.get(OtelAttr.REQUEST_MODEL, "unknown")
-            span = get_tracer().start_span(f"{operation} {span_name}")
-            span.set_attributes(attributes)
             if OBSERVABILITY_SETTINGS.SENSITIVE_DATA_ENABLED and messages:
                 _capture_messages(
                     span=span,
@@ -1319,6 +1301,24 @@ class ChatTelemetryLayer(Generic[OptionsCoT]):
             def _record_duration() -> None:
                 duration_state["duration"] = perf_counter() - start_time
 
+            try:
+                result_stream = cast(
+                    ResponseStream[ChatResponseUpdate, ChatResponse[Any]],
+                    super_get_response(
+                        messages=messages,
+                        stream=True,
+                        options=opts,
+                        compaction_strategy=compaction_strategy,
+                        tokenizer=tokenizer,
+                        function_invocation_kwargs=function_invocation_kwargs,
+                        client_kwargs=merged_client_kwargs,
+                    ),
+                )
+            except Exception as exception:
+                capture_exception(span=span, exception=exception, timestamp=time_ns())
+                _close_span()
+                raise
+
             async def _finalize_stream() -> None:
                 from ._types import ChatResponse
 
@@ -1357,11 +1357,18 @@ class ChatTelemetryLayer(Generic[OptionsCoT]):
                 finally:
                     _close_span()
 
-            # Register a weak reference callback to close the span if stream is garbage collected
-            # without being consumed. This ensures spans don't leak if users don't consume streams.
-            wrapped_stream: ResponseStream[ChatResponseUpdate, ChatResponse[Any]] = result_stream.with_cleanup_hook(
-                _record_duration
-            ).with_cleanup_hook(_finalize_stream)
+            # The pull context manager attaches the span around each underlying iterator pull so
+            # that child spans created during the pull (e.g. HTTP requests, inner tool execution)
+            # are parented under this chat span. Attach and detach happen in the same async
+            # context as the pull, avoiding cross-context cleanup issues. The weakref finalizer
+            # ensures the span is closed even if the stream is garbage collected without being
+            # consumed.
+            wrapped_stream: ResponseStream[ChatResponseUpdate, ChatResponse[Any]] = (
+                result_stream
+                .with_cleanup_hook(_record_duration)
+                .with_cleanup_hook(_finalize_stream)
+                .with_pull_context_manager(lambda: _activate_span(span))
+            )
             weakref.finalize(wrapped_stream, _close_span)
             return wrapped_stream
 
@@ -1543,23 +1550,8 @@ class AgentTelemetryLayer:
         inner_accumulated_usage_token = INNER_ACCUMULATED_USAGE.set({})
 
         if stream:
-            try:
-                run_result: object = execute()
-                if isinstance(run_result, ResponseStream):
-                    result_stream: ResponseStream[AgentResponseUpdate, AgentResponse[Any]] = run_result  # pyright: ignore[reportUnknownVariableType]
-                elif isinstance(run_result, Awaitable):
-                    result_stream = ResponseStream.from_awaitable(run_result)  # type: ignore[arg-type]  # pyright: ignore[reportArgumentType]
-                else:
-                    raise RuntimeError("Streaming telemetry requires a ResponseStream result.")
-            except Exception:
-                INNER_RESPONSE_TELEMETRY_CAPTURED_FIELDS.reset(inner_response_telemetry_captured_fields_token)
-                INNER_ACCUMULATED_USAGE.reset(inner_accumulated_usage_token)
-                raise
+            span = _start_streaming_span(attributes, OtelAttr.AGENT_NAME)
 
-            operation = attributes.get(OtelAttr.OPERATION, "operation")
-            span_name = attributes.get(OtelAttr.AGENT_NAME, "unknown")
-            span = get_tracer().start_span(f"{operation} {span_name}")
-            span.set_attributes(attributes)
             if OBSERVABILITY_SETTINGS.SENSITIVE_DATA_ENABLED and messages:
                 _capture_messages(
                     span=span,
@@ -1581,6 +1573,21 @@ class AgentTelemetryLayer:
             def _record_duration() -> None:
                 duration_state["duration"] = perf_counter() - start_time
 
+            try:
+                run_result: object = execute()
+                if isinstance(run_result, ResponseStream):
+                    result_stream: ResponseStream[AgentResponseUpdate, AgentResponse[Any]] = run_result  # pyright: ignore[reportUnknownVariableType]
+                elif isinstance(run_result, Awaitable):
+                    result_stream = ResponseStream.from_awaitable(run_result)  # type: ignore[arg-type]  # pyright: ignore[reportArgumentType]
+                else:
+                    raise RuntimeError("Streaming telemetry requires a ResponseStream result.")
+            except Exception as exception:
+                capture_exception(span=span, exception=exception, timestamp=time_ns())
+                INNER_RESPONSE_TELEMETRY_CAPTURED_FIELDS.reset(inner_response_telemetry_captured_fields_token)
+                INNER_ACCUMULATED_USAGE.reset(inner_accumulated_usage_token)
+                _close_span()
+                raise
+
             async def _finalize_stream() -> None:
                 from ._types import AgentResponse
 
@@ -1620,9 +1627,18 @@ class AgentTelemetryLayer:
                     INNER_ACCUMULATED_USAGE.reset(inner_accumulated_usage_token)
                     _close_span()
 
-            wrapped_stream: ResponseStream[AgentResponseUpdate, AgentResponse[Any]] = result_stream.with_cleanup_hook(
-                _record_duration
-            ).with_cleanup_hook(_finalize_stream)
+            # The pull context manager attaches the span around each underlying iterator pull so
+            # that child spans created during the pull (e.g. inner chat completion spans from the
+            # underlying ChatTelemetryLayer) are parented under this agent invoke span. Attach and
+            # detach happen in the same async context as the pull, avoiding cross-context cleanup
+            # issues. The weakref finalizer ensures the span is closed even if the stream is
+            # garbage collected without being consumed.
+            wrapped_stream: ResponseStream[AgentResponseUpdate, AgentResponse[Any]] = (
+                result_stream
+                .with_cleanup_hook(_record_duration)
+                .with_cleanup_hook(_finalize_stream)
+                .with_pull_context_manager(lambda: _activate_span(span))
+            )
             weakref.finalize(wrapped_stream, _close_span)
             return wrapped_stream
 
@@ -1809,6 +1825,27 @@ def get_function_span(
     )
 
 
+@contextlib.contextmanager
+def _activate_span(span: trace.Span) -> Generator[None]:
+    """Attach ``span`` as the current span in the OpenTelemetry context.
+
+    Designed to be used as a per-pull context manager registered on a
+    ``ResponseStream`` via ``with_pull_context_manager``: it attaches the span
+    before each underlying iterator pull and detaches immediately after, so
+    child spans created during the pull (HTTP clients, inner chat completions,
+    tool execution) are correctly parented under ``span``.
+
+    Because attach and detach happen within the same ``__anext__`` invocation
+    (and therefore the same async task / contextvars context), there is no risk
+    of "Failed to detach context" warnings from cross-context cleanup.
+    """
+    token = otel_context.attach(trace.set_span_in_context(span))
+    try:
+        yield
+    finally:
+        otel_context.detach(token)
+
+
 @contextlib.contextmanager
 def _get_span(
     attributes: dict[str, Any],
@@ -1831,6 +1868,29 @@ def _get_span(
         yield current_span
 
 
+def _start_streaming_span(attributes: dict[str, Any], span_name_attribute: str) -> trace.Span:
+    """Start a non-current span for a streaming operation.
+
+    Unlike :func:`_get_span`, the returned span is not attached to the current
+    OpenTelemetry context. The caller is responsible for:
+
+    - Ending the span via cleanup hooks on the wrapped
+      :class:`~agent_framework._types.ResponseStream`.
+    - Activating the span around each iterator pull via
+      :func:`_activate_span` registered with ``with_pull_context_manager`` so
+      that child spans created during stream production inherit it as parent.
+
+    Streaming spans are closed asynchronously in cleanup hooks that run in a
+    different async context than creation, so attaching the span at creation
+    time would cause "Failed to detach context" errors from OpenTelemetry.
+    """
+    operation = attributes.get(OtelAttr.OPERATION, "operation")
+    span_name = attributes.get(span_name_attribute, "unknown")
+    span = get_tracer().start_span(f"{operation} {span_name}")
+    span.set_attributes(attributes)
+    return span
+
+
 def _get_instructions_from_options(options: Any) -> str | list[str] | None:
     """Extract instructions from options dict."""
     if options is None:

{agent_framework_core-1.2.0 → agent_framework_core-1.2.2}/pyproject.toml

@@ -4,7 +4,7 @@ description = "Microsoft Agent Framework for building AI Agents with Python. Thi
 authors = [{ name = "Microsoft", email = "af-support@microsoft.com"}]
 readme = "README.md"
 requires-python = ">=3.10"
-version = "1.2.0"
+version = "1.2.2"
 license-files = ["LICENSE"]
 urls.homepage = "https://aka.ms/agent-framework"
 urls.source = "https://github.com/microsoft/agent-framework/tree/main/python"