Merge branch 'main' into local-branch-python-add-reset-to-workflow

python/packages/ag-ui/AGENTS.md

@@ -10,10 +10,12 @@ AG-UI protocol integration for building agent UIs with the AG-UI standard.
 - **`AGUIHttpService`** - HTTP service for AG-UI endpoints
 - **`AGUIEventConverter`** - Converts between Agent Framework and AG-UI events
 - **`add_agent_framework_fastapi_endpoint()`** - Add AG-UI endpoint to FastAPI app (`SupportsAgentRun` or `Workflow`)
+- **`InMemoryAGUIThreadSnapshotStore`** - Memory-only latest AG-UI Thread Snapshot store for local development, demos, and tests
 
 ## Types
 
 - **`AGUIRequest`** / **`AGUIChatOptions`** - Request types
+- **`AGUIThreadSnapshot`** / **`AGUIThreadSnapshotStore`** - Replayable thread snapshot model and scoped async store protocol
 - **`availableInterrupts` / `resume`** - Optional interrupt configuration and continuation payloads
 - **`AgentState`** / **`RunMetadata`** - State management types
 - **`PredictStateConfig`** - Configuration for state prediction

python/packages/ag-ui/README.md

@@ -198,6 +198,71 @@ The `dependencies` parameter accepts any FastAPI dependency, enabling integratio
 
 For a complete authentication example, see [getting_started/server.py](getting_started/server.py).
 
+## AG-UI Thread Snapshots
+
+AG-UI Thread Snapshot persistence is opt-in and disabled by default. Existing endpoints keep their current behavior
+unless you provide a `snapshot_store`.
+
+Thread snapshots let an AG-UI frontend recover replayable UI state after a refresh. When snapshot persistence is
+enabled, the endpoint stores the latest replayable snapshot for an AG-UI Thread within an application-defined
+Snapshot Scope. A Hydrate Request is an AG-UI request with a known `threadId`, `messages: []`, and no `resume`
+payload. Hydration replays the stored Shared State, message snapshot, and interruption metadata when available,
+then finishes without invoking the wrapped agent or workflow.
+
+Use the built-in in-memory store for local development, demos, and tests:
+
+```python
+from fastapi import FastAPI
+
+from agent_framework.ag_ui import InMemoryAGUIThreadSnapshotStore, add_agent_framework_fastapi_endpoint
+
+app = FastAPI()
+agent = ...
+snapshot_store = InMemoryAGUIThreadSnapshotStore(max_snapshots=500)
+
+
+def resolve_snapshot_scope(request):
+    # Local demo scope. Production apps should derive the scope from authenticated user or tenant context.
+    del request
+    return "local-demo"
+
+
+add_agent_framework_fastapi_endpoint(
+    app,
+    agent,
+    "/",
+    snapshot_store=snapshot_store,
+    snapshot_scope_resolver=resolve_snapshot_scope,
+)
+```
+
+A frontend can then hydrate the latest stored snapshot for the scoped thread:
+
+```json
+{
+  "threadId": "thread-1",
+  "messages": []
+}
+```
+
+Endpoint configuration requires `snapshot_scope_resolver` whenever a snapshot store is configured, including when
+the store is already set on a pre-wrapped `AgentFrameworkAgent` or `AgentFrameworkWorkflow`. The resolver returns
+the application-defined Snapshot Scope used with the AG-UI Thread id as the storage key.
+
+AG-UI Thread ids identify AG-UI Threads; they do not authorize snapshot access. Do not treat a thread id as a bearer
+credential or tenant boundary. Production applications must authenticate and authorize every AG-UI endpoint request
+and choose a Snapshot Scope that represents the app's real access boundary, such as an authenticated user, tenant,
+or workspace. Do not rely on untrusted client-provided fields by themselves to choose that boundary.
+
+Stored snapshots are untrusted application data with confidentiality impact. They may contain sensitive user text,
+model output, tool results, function arguments, UI payloads, Shared State, and interruption data. The built-in
+`InMemoryAGUIThreadSnapshotStore` is in-memory only, process-local, bounded, latest-only, and not durable production
+storage. It is cleared on process restart and is not shared across workers.
+
+No file-backed AG-UI snapshot store is provided by the package. Applications that need durable persistence should
+provide an app-owned implementation of the `AGUIThreadSnapshotStore` protocol and own storage hardening, including
+encryption, access control, retention, audit, data residency, and deletion behavior.
+
 ## Architecture
 
 The package uses a clean, orchestrator-based architecture:

python/packages/ag-ui/agent_framework_ag_ui/__init__.py

@@ -9,6 +9,15 @@ from ._client import AGUIChatClient
 from ._endpoint import add_agent_framework_fastapi_endpoint
 from ._event_converters import AGUIEventConverter
 from ._http_service import AGUIHttpService
+from ._snapshots import (
+    DEFAULT_MAX_THREAD_SNAPSHOTS,
+    AGUIThreadID,
+    AGUIThreadSnapshot,
+    AGUIThreadSnapshotStore,
+    InMemoryAGUIThreadSnapshotStore,
+    SnapshotScope,
+    SnapshotScopeResolver,
+)
 from ._state import state_update
 from ._types import AgentState, AGUIChatOptions, AGUIRequest, PredictStateConfig, RunMetadata
 from ._workflow import AgentFrameworkWorkflow, WorkflowFactory
@@ -31,9 +40,16 @@ __all__ = [
     "AGUIEventConverter",
     "AGUIHttpService",
     "AGUIRequest",
+    "AGUIThreadID",
+    "AGUIThreadSnapshot",
+    "AGUIThreadSnapshotStore",
     "AgentState",
+    "InMemoryAGUIThreadSnapshotStore",
     "PredictStateConfig",
     "RunMetadata",
+    "SnapshotScope",
+    "SnapshotScopeResolver",
+    "DEFAULT_MAX_THREAD_SNAPSHOTS",
     "DEFAULT_TAGS",
     "state_update",
     "__version__",

python/packages/ag-ui/agent_framework_ag_ui/_agent.py

@@ -10,6 +10,7 @@ from ag_ui.core import BaseEvent
 from agent_framework import SupportsAgentRun
 
 from ._agent_run import PendingApprovalEntry, run_agent_stream
+from ._snapshots import AGUIThreadSnapshotStore
 
 
 class AgentConfig:
@@ -21,6 +22,7 @@ class AgentConfig:
         predict_state_config: dict[str, dict[str, str]] | None = None,
         use_service_session: bool = False,
         require_confirmation: bool = True,
+        snapshot_store: AGUIThreadSnapshotStore | None = None,
     ):
         """Initialize agent configuration.
 
@@ -29,11 +31,14 @@ class AgentConfig:
             predict_state_config: Configuration for predictive state updates
             use_service_session: Whether the agent session is service-managed
             require_confirmation: Whether predictive updates require user confirmation before applying
+            snapshot_store: Optional AG-UI Thread Snapshot store. Snapshot persistence remains inactive unless
+                endpoint setup also provides an explicit Snapshot Scope resolver.
         """
         self.state_schema = self._normalize_state_schema(state_schema)
         self.predict_state_config = predict_state_config or {}
         self.use_service_session = use_service_session
         self.require_confirmation = require_confirmation
+        self.snapshot_store = snapshot_store
 
     @staticmethod
     def _normalize_state_schema(state_schema: Any | None) -> dict[str, Any]:
@@ -79,6 +84,7 @@ class AgentFrameworkAgent:
         predict_state_config: dict[str, dict[str, str]] | None = None,
         require_confirmation: bool = True,
         use_service_session: bool = False,
+        snapshot_store: AGUIThreadSnapshotStore | None = None,
     ):
         """Initialize the AG-UI compatible agent wrapper.
 
@@ -90,6 +96,8 @@ class AgentFrameworkAgent:
             predict_state_config: Configuration for predictive state updates
             require_confirmation: Whether predictive updates require user confirmation before applying
             use_service_session: Whether the agent session is service-managed
+            snapshot_store: Optional AG-UI Thread Snapshot store. Snapshot persistence remains inactive unless
+                endpoint setup also provides an explicit Snapshot Scope resolver.
         """
         self.agent = agent
         self.name = name or getattr(agent, "name", "agent")
@@ -100,6 +108,7 @@ class AgentFrameworkAgent:
             predict_state_config=predict_state_config,
             use_service_session=use_service_session,
             require_confirmation=require_confirmation,
+            snapshot_store=snapshot_store,
         )
 
         # Server-side registry of pending approval requests.
@@ -110,6 +119,11 @@ class AgentFrameworkAgent:
         self._pending_approvals: OrderedDict[str, PendingApprovalEntry] = OrderedDict()
         self._pending_approvals_max_size: int = 10_000
 
+    @property
+    def snapshot_store(self) -> AGUIThreadSnapshotStore | None:
+        """Configured AG-UI Thread Snapshot store, if any."""
+        return self.config.snapshot_store
+
     async def run(
         self,
         input_data: dict[str, Any],

python/packages/ag-ui/agent_framework_ag_ui/_agent_run.py

@@ -4,6 +4,7 @@
 
 from __future__ import annotations  # noqa: I001
 
+import copy
 import json
 import logging
 import uuid
@@ -52,9 +53,11 @@ from ._run_common import (
     _extract_tool_result_display,  # type: ignore
     _has_only_tool_calls,  # type: ignore
     _normalize_resume_interrupts,  # type: ignore
+    _reconstruct_messages_from_thread_snapshot,  # type: ignore
     _resolve_ui_payload,  # type: ignore
     _stringify_tool_result,  # type: ignore
 )
+from ._snapshots import AGUIThreadSnapshot, _DEFAULT_STATE_INPUT_KEY, _SNAPSHOT_SCOPE_INPUT_KEY
 from ._utils import (
     canonical_function_arguments,
     convert_agui_tools_to_agent_framework,
@@ -748,6 +751,85 @@ def _build_messages_snapshot(
     return MessagesSnapshotEvent(messages=all_messages)  # type: ignore[arg-type]
 
 
+def _event_messages_to_snapshot_dicts(messages: list[Any]) -> list[dict[str, Any]]:
+    """Convert AG-UI message event models back to plain snapshot dictionaries."""
+    safe_messages = make_json_safe(messages)
+    if not isinstance(safe_messages, list):
+        return []
+    return [cast(dict[str, Any], message) for message in safe_messages if isinstance(message, dict)]
+
+
+def _text_events_to_snapshot_messages(events: list[BaseEvent]) -> list[dict[str, Any]]:
+    """Convert streamed text-message events into snapshot message dictionaries."""
+    messages: list[dict[str, Any]] = []
+    messages_by_id: dict[str, dict[str, Any]] = {}
+    for event in events:
+        if isinstance(event, TextMessageStartEvent):
+            message: dict[str, Any] = {"id": event.message_id, "role": event.role, "content": ""}
+            messages.append(message)
+            messages_by_id[event.message_id] = message
+        elif isinstance(event, TextMessageContentEvent):
+            open_message = messages_by_id.get(event.message_id)
+            if open_message is not None:
+                open_message["content"] = f"{open_message['content']}{event.delta}"
+    return [message for message in messages if message.get("content")]
+
+
+async def _hydrate_thread_snapshot(
+    *,
+    config: AgentConfig,
+    scope: str,
+    thread_id: str,
+    run_id: str,
+) -> AsyncGenerator[BaseEvent]:
+    """Replay the latest stored AG-UI Thread Snapshot without invoking the agent."""
+    yield RunStartedEvent(run_id=run_id, thread_id=thread_id)
+    if config.snapshot_store is None:
+        yield _build_run_finished_event(run_id=run_id, thread_id=thread_id)
+        return
+
+    snapshot = await config.snapshot_store.get(scope=scope, thread_id=thread_id)
+    if snapshot is None:
+        yield _build_run_finished_event(run_id=run_id, thread_id=thread_id)
+        return
+
+    if snapshot.state is not None:
+        yield StateSnapshotEvent(snapshot=snapshot.state)
+    if snapshot.messages:
+        yield MessagesSnapshotEvent(messages=snapshot.messages)  # type: ignore[arg-type]
+    yield _build_run_finished_event(run_id=run_id, thread_id=thread_id, interrupts=snapshot.interrupt)
+
+
+async def _save_thread_snapshot(
+    *,
+    config: AgentConfig,
+    scope: str | None,
+    thread_id: str,
+    messages: list[dict[str, Any]],
+    state: dict[str, Any] | None,
+    interrupt: list[dict[str, Any]] | None,
+) -> None:
+    """Save the latest replayable AG-UI Thread Snapshot when persistence is configured."""
+    if config.snapshot_store is None or scope is None:
+        return
+
+    try:
+        await config.snapshot_store.save(
+            scope=scope,
+            thread_id=thread_id,
+            snapshot=AGUIThreadSnapshot(messages=messages, state=state, interrupt=interrupt),
+        )
+    except Exception:
+        # The run itself already streamed successfully; a transient store failure
+        # must not surface as RUN_ERROR for a completed run. The previous snapshot
+        # stays available for hydration.
+        logger.exception(
+            "Failed to save AG-UI Thread Snapshot for scope=%s thread_id=%s; keeping previous snapshot.",
+            scope,
+            thread_id,
+        )
+
+
 async def run_agent_stream(
     input_data: dict[str, Any],
     agent: SupportsAgentRun,
@@ -774,15 +856,53 @@ async def run_agent_stream(
     # Parse IDs
     thread_id = input_data.get("thread_id") or input_data.get("threadId") or str(uuid.uuid4())
     run_id = input_data.get("run_id") or input_data.get("runId") or str(uuid.uuid4())
-
-    # Initialize flow state with schema defaults
-    flow = FlowState()
-    if input_data.get("state"):
-        flow.current_state = dict(input_data["state"])
+    snapshot_scope = cast(str | None, input_data.get(_SNAPSHOT_SCOPE_INPUT_KEY))
 
     state_schema = cast(dict[str, Any], getattr(config, "state_schema", {}) or {})
     predict_state_config = cast(dict[str, dict[str, str]], getattr(config, "predict_state_config", {}) or {})
 
+    # Normalize messages
+    available_interrupts = input_data.get("available_interrupts") or input_data.get("availableInterrupts")
+    raw_messages: list[dict[str, Any]] = input_data.get("messages", []) or []
+    resume_payload = _extract_resume_payload(input_data)
+    if config.snapshot_store is not None and snapshot_scope is not None and not raw_messages and resume_payload is None:
+        async for event in _hydrate_thread_snapshot(
+            config=config,
+            scope=snapshot_scope,
+            thread_id=thread_id,
+            run_id=run_id,
+        ):
+            yield event
+        return
+
+    stored_snapshot: AGUIThreadSnapshot | None = None
+    if config.snapshot_store is not None and snapshot_scope is not None:
+        stored_snapshot = await config.snapshot_store.get(scope=snapshot_scope, thread_id=thread_id)
+        if stored_snapshot is not None and resume_payload is None:
+            raw_messages = _reconstruct_messages_from_thread_snapshot(
+                stored_messages=stored_snapshot.messages,
+                incoming_messages=raw_messages,
+                stored_interrupt=stored_snapshot.interrupt,
+            )
+
+    # Initialize flow state with stored state plus request-provided overrides.
+    flow = FlowState()
+    request_state = input_data.get("state")
+    if stored_snapshot is not None and stored_snapshot.state is not None:
+        flow.current_state = dict(stored_snapshot.state)
+        if isinstance(request_state, dict):
+            flow.current_state.update(request_state)
+    elif isinstance(request_state, dict):
+        flow.current_state = dict(request_state)
+
+    # Apply endpoint-deferred defaults only for keys missing from both the stored
+    # snapshot state and the request state, so defaults never reset persisted state.
+    deferred_default_state = cast(dict[str, Any] | None, input_data.get(_DEFAULT_STATE_INPUT_KEY))
+    if deferred_default_state:
+        for key, value in deferred_default_state.items():
+            if key not in flow.current_state:
+                flow.current_state[key] = copy.deepcopy(value)
+
     # Apply schema defaults for missing state keys
     if state_schema:
         for key, schema in state_schema.items():
@@ -801,10 +921,7 @@ async def run_agent_stream(
             current_state=flow.current_state,
         )
 
-    # Normalize messages
-    available_interrupts = input_data.get("available_interrupts") or input_data.get("availableInterrupts")
-    raw_messages = list(cast(list[dict[str, Any]], input_data.get("messages", []) or []))
-    resume_messages = _resume_to_tool_messages(_extract_resume_payload(input_data))
+    resume_messages = _resume_to_tool_messages(resume_payload)
     if available_interrupts:
         logger.debug("Received available interrupts metadata: %s", available_interrupts)
     if resume_messages:
@@ -892,8 +1009,24 @@ async def run_agent_stream(
         # Emit approved state snapshot before confirmation message
         if approved_state_snapshot_emitted:
             yield StateSnapshotEvent(snapshot=flow.current_state)
-        for event in _handle_step_based_approval(messages):
+        confirmation_events = _handle_step_based_approval(messages)
+        for event in confirmation_events:
             yield event
+        # Persist the completed confirmation turn with interrupt=None so hydration
+        # does not replay the stale pending interrupt after the user responded.
+        persisted_messages = snapshot_messages + _text_events_to_snapshot_messages(confirmation_events)
+        if resume_payload is not None and stored_snapshot is not None:
+            # Resume requests carry only the synthesized interrupt response, so prepend
+            # the stored thread history to avoid persisting a truncated thread.
+            persisted_messages = [copy.deepcopy(message) for message in stored_snapshot.messages] + persisted_messages
+        await _save_thread_snapshot(
+            config=config,
+            scope=snapshot_scope,
+            thread_id=thread_id,
+            messages=persisted_messages,
+            state=cast(dict[str, Any], make_json_safe(flow.current_state)) if flow.current_state else None,
+            interrupt=None,
+        )
         yield _build_run_finished_event(run_id=run_id, thread_id=thread_id)
         return
 
@@ -905,6 +1038,9 @@ async def run_agent_stream(
     # Stream from agent - emit RunStarted after first update to get service IDs
     run_started_emitted = False
     all_updates: list[Any] = []  # Collect for structured output processing
+    latest_state_snapshot: dict[str, Any] | None = (
+        cast(dict[str, Any], make_json_safe(flow.current_state)) if flow.current_state else None
+    )
     response_stream = agent.run(messages, stream=True, **run_kwargs)
     stream = await _normalize_response_stream(response_stream)
     async for update in stream:
@@ -934,6 +1070,7 @@ async def run_agent_stream(
                 yield CustomEvent(name="PredictState", value=predict_state_value)
             # Emit initial state snapshot only if we have both state_schema and state
             if state_schema and flow.current_state:
+                latest_state_snapshot = cast(dict[str, Any], make_json_safe(flow.current_state))
                 yield StateSnapshotEvent(snapshot=flow.current_state)
             run_started_emitted = True
 
@@ -975,6 +1112,8 @@ async def run_agent_stream(
                 skip_text,
                 config.require_confirmation,
             ):
+                if isinstance(event, StateSnapshotEvent):
+                    latest_state_snapshot = cast(dict[str, Any], make_json_safe(event.snapshot))
                 yield event
 
         # Stop if waiting for approval
@@ -1019,6 +1158,7 @@ async def run_agent_stream(
 
                 if state_updates:
                     flow.current_state.update(state_updates)
+                    latest_state_snapshot = cast(dict[str, Any], make_json_safe(flow.current_state))
                     yield StateSnapshotEvent(snapshot=flow.current_state)
                     logger.info(f"Emitted StateSnapshotEvent with updates: {list(state_updates.keys())}")
 
@@ -1056,6 +1196,7 @@ async def run_agent_stream(
                                 if result:
                                     state_key, state_value = result
                                     flow.current_state[state_key] = state_value
+                                    latest_state_snapshot = cast(dict[str, Any], make_json_safe(flow.current_state))
                                     yield StateSnapshotEvent(snapshot=flow.current_state)
                             except json.JSONDecodeError:
                                 # Ignore malformed JSON in tool arguments for predictive state;
@@ -1136,7 +1277,12 @@ async def run_agent_stream(
     should_emit_snapshot = (
         flow.pending_tool_calls or flow.tool_results or flow.accumulated_text or flow.reasoning_messages
     )
+    latest_messages_snapshot = snapshot_messages
     if should_emit_snapshot:
+        # Always fold this turn's output into the persisted snapshot, even when the
+        # outbound MESSAGES_SNAPSHOT event is suppressed for predictive tools.
+        snapshot_event = _build_messages_snapshot(flow, snapshot_messages)
+        latest_messages_snapshot = _event_messages_to_snapshot_dicts(list(snapshot_event.messages))
         # Check if we should suppress for predictive tool
         last_tool_name = None
         if flow.tool_results:
@@ -1146,8 +1292,21 @@ async def run_agent_stream(
         if not _should_suppress_intermediate_snapshot(
             last_tool_name, predict_state_config, config.require_confirmation
         ):
-            yield _build_messages_snapshot(flow, snapshot_messages)
+            yield snapshot_event
 
     # Always emit RunFinished - confirm_changes tool call is complete (Start -> Args -> End)
     # The UI will show confirmation dialog and send a new request when user responds
+    persisted_messages = latest_messages_snapshot
+    if resume_payload is not None and stored_snapshot is not None:
+        # Resume requests carry only the synthesized interrupt response, so prepend
+        # the stored thread history to avoid persisting a truncated thread.
+        persisted_messages = [copy.deepcopy(message) for message in stored_snapshot.messages] + persisted_messages
+    await _save_thread_snapshot(
+        config=config,
+        scope=snapshot_scope,
+        thread_id=thread_id,
+        messages=persisted_messages,
+        state=latest_state_snapshot,
+        interrupt=flow.interrupts or None,
+    )
     yield _build_run_finished_event(run_id=run_id, thread_id=thread_id, interrupts=flow.interrupts)

python/packages/ag-ui/agent_framework_ag_ui/_endpoint.py

@@ -7,6 +7,7 @@ from __future__ import annotations
 import copy
 import logging
 from collections.abc import AsyncGenerator, Sequence
+from inspect import isawaitable
 from typing import Any
 
 from ag_ui.core import RunErrorEvent
@@ -17,12 +18,58 @@ from fastapi.params import Depends
 from fastapi.responses import StreamingResponse
 
 from ._agent import AgentFrameworkAgent
+from ._snapshots import (
+    _DEFAULT_STATE_INPUT_KEY,
+    _SNAPSHOT_SCOPE_INPUT_KEY,
+    AGUIThreadSnapshotStore,
+    SnapshotScopeResolver,
+)
 from ._types import AGUIRequest
 from ._workflow import AgentFrameworkWorkflow
 
 logger = logging.getLogger(__name__)
 
 
+def _get_snapshot_store(
+    protocol_runner: AgentFrameworkAgent | AgentFrameworkWorkflow,
+) -> AGUIThreadSnapshotStore | None:
+    if isinstance(protocol_runner, AgentFrameworkAgent):
+        return protocol_runner.config.snapshot_store
+    return protocol_runner.snapshot_store
+
+
+def _set_snapshot_store(
+    protocol_runner: AgentFrameworkAgent | AgentFrameworkWorkflow,
+    snapshot_store: AGUIThreadSnapshotStore,
+) -> None:
+    if isinstance(protocol_runner, AgentFrameworkAgent):
+        protocol_runner.config.snapshot_store = snapshot_store
+        return
+    protocol_runner.snapshot_store = snapshot_store
+
+
+def _configure_snapshot_persistence(
+    protocol_runner: AgentFrameworkAgent | AgentFrameworkWorkflow,
+    *,
+    snapshot_store: AGUIThreadSnapshotStore | None,
+    snapshot_scope_resolver: SnapshotScopeResolver | None,
+) -> None:
+    existing_snapshot_store = _get_snapshot_store(protocol_runner)
+    if snapshot_store is not None:
+        if existing_snapshot_store is not None and existing_snapshot_store is not snapshot_store:
+            raise ValueError("snapshot_store is already configured on the AG-UI runner.")
+        if existing_snapshot_store is None:
+            _set_snapshot_store(protocol_runner, snapshot_store)
+        existing_snapshot_store = snapshot_store
+
+    if existing_snapshot_store is not None and snapshot_scope_resolver is None:
+        raise ValueError(
+            "snapshot_scope_resolver is required when snapshot_store is configured. "
+            "AG-UI Thread ids identify threads but do not authorize snapshot access; "
+            "provide a resolver that returns an explicit Snapshot Scope."
+        )
+
+
 def add_agent_framework_fastapi_endpoint(
     app: FastAPI,
     agent: SupportsAgentRun | AgentFrameworkAgent | Workflow | AgentFrameworkWorkflow,
@@ -33,6 +80,8 @@ def add_agent_framework_fastapi_endpoint(
     default_state: dict[str, Any] | None = None,
     tags: list[str] | None = None,
     dependencies: Sequence[Depends] | None = None,
+    snapshot_store: AGUIThreadSnapshotStore | None = None,
+    snapshot_scope_resolver: SnapshotScopeResolver | None = None,
 ) -> None:
     """Add an AG-UI endpoint to a FastAPI app.
 
@@ -50,6 +99,10 @@ def add_agent_framework_fastapi_endpoint(
             These dependencies run before the endpoint handler. Use this to add
             authentication checks, rate limiting, or other middleware-like behavior.
             Example: `dependencies=[Depends(verify_api_key)]`
+        snapshot_store: Optional AG-UI Thread Snapshot store. Snapshot persistence is opt-in and requires an
+            explicit Snapshot Scope resolver.
+        snapshot_scope_resolver: Optional resolver for the application-defined Snapshot Scope. Required whenever
+            a snapshot store is configured because an AG-UI Thread id is not an authorization boundary.
     """
     protocol_runner: AgentFrameworkAgent | AgentFrameworkWorkflow
     if isinstance(agent, AgentFrameworkWorkflow):
@@ -63,10 +116,17 @@ def add_agent_framework_fastapi_endpoint(
             agent=agent,
             state_schema=state_schema,
             predict_state_config=predict_state_config,
+            snapshot_store=snapshot_store,
         )
     else:
         raise TypeError("agent must be SupportsAgentRun, Workflow, AgentFrameworkAgent, or AgentFrameworkWorkflow.")
 
+    _configure_snapshot_persistence(
+        protocol_runner,
+        snapshot_store=snapshot_store,
+        snapshot_scope_resolver=snapshot_scope_resolver,
+    )
+
     @app.post(path, tags=tags or ["AG-UI"], dependencies=dependencies, response_model=None)  # type: ignore[arg-type]
     async def agent_endpoint(request_body: AGUIRequest) -> StreamingResponse:
         """Handle AG-UI agent requests.
@@ -76,11 +136,23 @@ def add_agent_framework_fastapi_endpoint(
         """
         try:
             input_data = request_body.model_dump(exclude_none=True)
+            snapshot_persistence_active = False
+            if snapshot_scope_resolver is not None and _get_snapshot_store(protocol_runner) is not None:
+                snapshot_scope = snapshot_scope_resolver(request_body)
+                if isawaitable(snapshot_scope):
+                    snapshot_scope = await snapshot_scope
+                input_data[_SNAPSHOT_SCOPE_INPUT_KEY] = snapshot_scope
+                snapshot_persistence_active = True
             if default_state:
-                state = input_data.setdefault("state", {})
-                for key, value in default_state.items():
-                    if key not in state:
-                        state[key] = copy.deepcopy(value)
+                if snapshot_persistence_active:
+                    # Defer default application to the runner so defaults only fill keys
+                    # missing from both the stored snapshot state and the request state.
+                    input_data[_DEFAULT_STATE_INPUT_KEY] = copy.deepcopy(default_state)
+                else:
+                    state = input_data.setdefault("state", {})
+                    for key, value in default_state.items():
+                        if key not in state:
+                            state[key] = copy.deepcopy(value)
             logger.debug(
                 f"[{path}] Received request - Run ID: {input_data.get('run_id', 'no-run-id')}, "
                 f"Thread ID: {input_data.get('thread_id', 'no-thread-id')}, "

python/packages/ag-ui/agent_framework_ag_ui/_run_common.py

@@ -4,6 +4,7 @@
 
 from __future__ import annotations
 
+import copy
 import json
 import logging
 from collections.abc import Mapping
@@ -33,7 +34,7 @@ from agent_framework import Content
 
 from ._orchestration._predictive_state import PredictiveStateHandler
 from ._state import TOOL_RESULT_DISPLAY_KEY, TOOL_RESULT_STATE_KEY
-from ._utils import generate_event_id, make_json_safe
+from ._utils import generate_event_id, make_json_safe, normalize_agui_role
 
 logger = logging.getLogger(__name__)
 
@@ -733,3 +734,117 @@ def _emit_content(
         return _emit_text_reasoning(content, flow)
     logger.debug("Skipping unsupported content type in AG-UI emitter: %s", content_type)
     return events
+
+
+def _canonical_snapshot_message(message: dict[str, Any]) -> dict[str, Any]:
+    """Normalize an AG-UI message for identity comparison without generated ids."""
+    from ._message_adapters import agui_messages_to_snapshot_format
+
+    normalized_message = agui_messages_to_snapshot_format([copy.deepcopy(message)])[0]
+    normalized_message.pop("id", None)
+    return cast(dict[str, Any], make_json_safe(normalized_message))
+
+
+def _snapshot_messages_match(stored_message: dict[str, Any], incoming_message: dict[str, Any]) -> bool:
+    """Return whether an incoming message already represents the stored snapshot message."""
+    stored_id = stored_message.get("id")
+    incoming_id = incoming_message.get("id")
+    if stored_id and incoming_id:
+        return str(stored_id) == str(incoming_id)
+    return _canonical_snapshot_message(stored_message) == _canonical_snapshot_message(incoming_message)
+
+
+def _latest_user_message_index(messages: list[dict[str, Any]]) -> int | None:
+    """Find the newest incoming user message index."""
+    for index in range(len(messages) - 1, -1, -1):
+        if normalize_agui_role(messages[index].get("role", "user")) == "user":
+            return index
+    return None
+
+
+def _known_tool_call_ids(
+    stored_messages: list[dict[str, Any]],
+    stored_interrupt: list[dict[str, Any]] | None,
+) -> set[str]:
+    """Collect tool call ids the backend previously issued for this thread."""
+    known_ids: set[str] = set()
+    for message in stored_messages:
+        tool_calls = message.get("tool_calls") or message.get("toolCalls") or []
+        if not isinstance(tool_calls, list):
+            continue
+        for tool_call in cast(list[Any], tool_calls):
+            if isinstance(tool_call, dict):
+                tool_call_id = cast(dict[str, Any], tool_call).get("id")
+                if tool_call_id:
+                    known_ids.add(str(tool_call_id))
+    for interrupt in stored_interrupt or []:
+        interrupt_id = interrupt.get("id")
+        if interrupt_id:
+            known_ids.add(str(interrupt_id))
+    return known_ids
+
+
+def _filter_untrusted_suffix(
+    incoming_suffix: list[dict[str, Any]],
+    *,
+    stored_messages: list[dict[str, Any]],
+    stored_interrupt: list[dict[str, Any]] | None,
+) -> list[dict[str, Any]]:
+    """Drop client-forged non-user messages before promoting them to stored history.
+
+    Only the user's own turns and tool results answering backend-issued tool calls
+    (including pending interrupts) may extend the authoritative thread history.
+    """
+    known_ids: set[str] | None = None
+    filtered: list[dict[str, Any]] = []
+    for message in incoming_suffix:
+        raw_role = str(message.get("role", "")).lower()
+        if raw_role == "user":
+            filtered.append(message)
+            continue
+        if raw_role == "tool":
+            tool_call_id = message.get("toolCallId") or message.get("tool_call_id") or message.get("actionExecutionId")
+            if known_ids is None:
+                known_ids = _known_tool_call_ids(stored_messages, stored_interrupt)
+            if tool_call_id and str(tool_call_id) in known_ids:
+                filtered.append(message)
+                continue
+        logger.warning(
+            "Dropping client-supplied %r message from the incoming thread suffix; "
+            "only user turns and tool results for backend-issued tool calls extend stored history.",
+            raw_role or "unknown",
+        )
+    return filtered
+
+
+def _reconstruct_messages_from_thread_snapshot(
+    *,
+    stored_messages: list[dict[str, Any]],
+    incoming_messages: list[dict[str, Any]],
+    stored_interrupt: list[dict[str, Any]] | None = None,
+) -> list[dict[str, Any]]:
+    """Combine backend-owned prior history with the request-owned new user turn."""
+    if not stored_messages or not incoming_messages:
+        return incoming_messages
+
+    incoming_suffix: list[dict[str, Any]]
+    if len(incoming_messages) >= len(stored_messages) and all(
+        _snapshot_messages_match(stored_message, incoming_message)
+        for stored_message, incoming_message in zip(stored_messages, incoming_messages)
+    ):
+        incoming_suffix = incoming_messages[len(stored_messages) :]
+    else:
+        latest_user_index = _latest_user_message_index(incoming_messages)
+        if latest_user_index is None:
+            return incoming_messages
+        incoming_suffix = incoming_messages[latest_user_index:]
+
+    incoming_suffix = _filter_untrusted_suffix(
+        incoming_suffix,
+        stored_messages=stored_messages,
+        stored_interrupt=stored_interrupt,
+    )
+
+    return [copy.deepcopy(message) for message in stored_messages] + [
+        copy.deepcopy(message) for message in incoming_suffix
+    ]

python/packages/ag-ui/agent_framework_ag_ui/_snapshots.py

@@ -0,0 +1,202 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""AG-UI Thread Snapshot storage primitives."""
+
+from __future__ import annotations
+
+import copy
+from collections.abc import Awaitable, Callable
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Any, Protocol, TypeAlias, runtime_checkable
+
+if TYPE_CHECKING:
+    from ._types import AGUIRequest
+
+SnapshotScope: TypeAlias = str
+"""Application-defined scope for authorizing access to AG-UI Thread Snapshots."""
+
+AGUIThreadID: TypeAlias = str
+"""AG-UI Thread identifier within a Snapshot Scope."""
+
+SnapshotScopeResolver: TypeAlias = Callable[["AGUIRequest"], str | Awaitable[str]]
+"""Callable that resolves the Snapshot Scope for an AG-UI endpoint request."""
+
+_SnapshotKey: TypeAlias = tuple[SnapshotScope, AGUIThreadID]
+
+DEFAULT_MAX_THREAD_SNAPSHOTS = 1_000
+_SNAPSHOT_SCOPE_INPUT_KEY = "__ag_ui_snapshot_scope"
+_DEFAULT_STATE_INPUT_KEY = "__ag_ui_default_state"
+
+
+@dataclass(slots=True)
+class AGUIThreadSnapshot:
+    """Replayable AG-UI Thread state.
+
+    AG-UI Thread Snapshots intentionally contain only data that can be replayed
+    to a UI: message snapshots, optional Shared State, and optional interruption
+    state. They do not include raw events, request metadata, auth claims,
+    diagnostics, traces, or provider responses.
+
+    Attributes:
+        messages: Replayable AG-UI message snapshots.
+        state: Optional AG-UI Shared State snapshot.
+        interrupt: Optional interruption state from ``RUN_FINISHED.interrupt``.
+    """
+
+    messages: list[dict[str, Any]] = field(default_factory=list)
+    state: dict[str, Any] | None = None
+    interrupt: list[dict[str, Any]] | None = None
+
+
+@runtime_checkable
+class AGUIThreadSnapshotStore(Protocol):
+    """Async store for latest AG-UI Thread Snapshots keyed by scope and thread id."""
+
+    async def save(
+        self,
+        *,
+        scope: SnapshotScope,
+        thread_id: AGUIThreadID,
+        snapshot: AGUIThreadSnapshot,
+    ) -> None:
+        """Save the latest snapshot for an AG-UI Thread within a Snapshot Scope.
+
+        Args:
+            scope: Application-defined Snapshot Scope. This is part of the
+                storage key and must represent the app's authorization boundary.
+            thread_id: AG-UI Thread id within the scope.
+            snapshot: Snapshot to save.
+        """
+        ...
+
+    async def get(
+        self,
+        *,
+        scope: SnapshotScope,
+        thread_id: AGUIThreadID,
+    ) -> AGUIThreadSnapshot | None:
+        """Get the latest snapshot for an AG-UI Thread within a Snapshot Scope.
+
+        Args:
+            scope: Application-defined Snapshot Scope.
+            thread_id: AG-UI Thread id within the scope.
+
+        Returns:
+            The latest snapshot, or ``None`` when no snapshot exists for the key.
+        """
+        ...
+
+    async def delete(
+        self,
+        *,
+        scope: SnapshotScope,
+        thread_id: AGUIThreadID,
+    ) -> bool:
+        """Delete the latest snapshot for an AG-UI Thread within a Snapshot Scope.
+
+        Args:
+            scope: Application-defined Snapshot Scope.
+            thread_id: AG-UI Thread id within the scope.
+
+        Returns:
+            ``True`` when a snapshot was deleted, otherwise ``False``.
+        """
+        ...
+
+    async def clear(self, *, scope: SnapshotScope | None = None) -> None:
+        """Clear saved snapshots.
+
+        Args:
+            scope: Optional Snapshot Scope to clear. When omitted, all in-memory
+                snapshots are cleared.
+        """
+        ...
+
+
+class InMemoryAGUIThreadSnapshotStore:
+    """Bounded memory-only latest snapshot store for local development, demos, and tests.
+
+    This store keeps at most one snapshot per ``(scope, thread_id)`` key. It is
+    process-local and not durable production storage.
+    """
+
+    def __init__(self, *, max_snapshots: int = DEFAULT_MAX_THREAD_SNAPSHOTS) -> None:
+        """Initialize the in-memory snapshot store.
+
+        Keyword Args:
+            max_snapshots: Maximum number of scoped thread snapshots to retain.
+
+        Raises:
+            ValueError: If ``max_snapshots`` is less than 1.
+        """
+        if max_snapshots < 1:
+            raise ValueError("max_snapshots must be greater than 0.")
+        self._max_snapshots = max_snapshots
+        self._snapshots: dict[_SnapshotKey, AGUIThreadSnapshot] = {}
+
+    async def save(
+        self,
+        *,
+        scope: SnapshotScope,
+        thread_id: AGUIThreadID,
+        snapshot: AGUIThreadSnapshot,
+    ) -> None:
+        """Save the latest snapshot for an AG-UI Thread within a Snapshot Scope."""
+        key = self._key(scope=scope, thread_id=thread_id)
+        if key in self._snapshots:
+            del self._snapshots[key]
+        self._snapshots[key] = copy.deepcopy(snapshot)
+        self._evict_oldest()
+
+    async def get(
+        self,
+        *,
+        scope: SnapshotScope,
+        thread_id: AGUIThreadID,
+    ) -> AGUIThreadSnapshot | None:
+        """Get the latest snapshot for an AG-UI Thread within a Snapshot Scope."""
+        snapshot = self._snapshots.get(self._key(scope=scope, thread_id=thread_id))
+        return copy.deepcopy(snapshot) if snapshot is not None else None
+
+    async def delete(
+        self,
+        *,
+        scope: SnapshotScope,
+        thread_id: AGUIThreadID,
+    ) -> bool:
+        """Delete the latest snapshot for an AG-UI Thread within a Snapshot Scope."""
+        key = self._key(scope=scope, thread_id=thread_id)
+        if key not in self._snapshots:
+            return False
+        del self._snapshots[key]
+        return True
+
+    async def clear(self, *, scope: SnapshotScope | None = None) -> None:
+        """Clear saved snapshots, optionally limited to one Snapshot Scope."""
+        if scope is None:
+            self._snapshots.clear()
+            return
+
+        normalized_scope = self._normalize_key_part(scope, "scope")
+        for key in list(self._snapshots):
+            if key[0] == normalized_scope:
+                del self._snapshots[key]
+
+    @classmethod
+    def _key(cls, *, scope: SnapshotScope, thread_id: AGUIThreadID) -> _SnapshotKey:
+        return (
+            cls._normalize_key_part(scope, "scope"),
+            cls._normalize_key_part(thread_id, "thread_id"),
+        )
+
+    @staticmethod
+    def _normalize_key_part(value: str, name: str) -> str:
+        if not isinstance(value, str):
+            raise TypeError(f"{name} must be a string.")
+        if not value:
+            raise ValueError(f"{name} must be a non-empty string.")
+        return value
+
+    def _evict_oldest(self) -> None:
+        while len(self._snapshots) > self._max_snapshots:
+            del self._snapshots[next(iter(self._snapshots))]

python/packages/ag-ui/agent_framework_ag_ui/_workflow.py

@@ -4,18 +4,203 @@
 
 from __future__ import annotations
 
+import copy
+import logging
 import uuid
 from collections.abc import AsyncGenerator, Callable
-from typing import Any
+from typing import Any, cast
 
-from ag_ui.core import BaseEvent
+from ag_ui.core import (
+    BaseEvent,
+    MessagesSnapshotEvent,
+    RunErrorEvent,
+    RunFinishedEvent,
+    RunStartedEvent,
+    StateSnapshotEvent,
+    TextMessageContentEvent,
+    TextMessageEndEvent,
+    TextMessageStartEvent,
+    ToolCallArgsEvent,
+    ToolCallResultEvent,
+    ToolCallStartEvent,
+)
 from agent_framework import Workflow
 
+from ._message_adapters import agui_messages_to_snapshot_format
+from ._run_common import (
+    _build_run_finished_event,
+    _extract_resume_payload,
+    _reconstruct_messages_from_thread_snapshot,
+)
+from ._snapshots import (
+    _DEFAULT_STATE_INPUT_KEY,
+    _SNAPSHOT_SCOPE_INPUT_KEY,
+    AGUIThreadSnapshot,
+    AGUIThreadSnapshotStore,
+)
+from ._utils import generate_event_id, make_json_safe
 from ._workflow_run import run_workflow_stream
 
+logger = logging.getLogger(__name__)
+
 WorkflowFactory = Callable[[str], Workflow]
 
 
+def _event_messages_to_snapshot_dicts(messages: list[Any]) -> list[dict[str, Any]]:
+    """Convert AG-UI message event models to plain snapshot dictionaries."""
+    safe_messages = make_json_safe(messages)
+    if not isinstance(safe_messages, list):
+        return []
+    return [cast(dict[str, Any], message) for message in safe_messages if isinstance(message, dict)]
+
+
+class _WorkflowSnapshotBuilder:
+    """Capture replayable workflow protocol output without retaining raw events."""
+
+    def __init__(self, raw_messages: list[dict[str, Any]]) -> None:
+        self._synthesized_messages = agui_messages_to_snapshot_format(raw_messages)
+        self._emitted_messages: list[dict[str, Any]] | None = None
+        self._open_text_message: dict[str, Any] | None = None
+        self._tool_call_message: dict[str, Any] | None = None
+        self._tool_calls_by_id: dict[str, dict[str, Any]] = {}
+        self.state: dict[str, Any] | None = None
+        self.interrupt: list[dict[str, Any]] | None = None
+
+    def observe(self, event: BaseEvent) -> None:
+        """Fold one replayable AG-UI event into the latest snapshot state."""
+        if isinstance(event, StateSnapshotEvent):
+            state = make_json_safe(event.snapshot)
+            if isinstance(state, dict):
+                self.state = cast(dict[str, Any], state)
+            return
+
+        if isinstance(event, MessagesSnapshotEvent):
+            self._emitted_messages = _event_messages_to_snapshot_dicts(list(event.messages))
+            return
+
+        if isinstance(event, RunFinishedEvent):
+            interrupt = make_json_safe(getattr(event, "interrupt", None))
+            if isinstance(interrupt, list):
+                self.interrupt = [cast(dict[str, Any], item) for item in interrupt if isinstance(item, dict)]
+            return
+
+        if self._emitted_messages is not None:
+            return
+
+        if isinstance(event, TextMessageStartEvent):
+            self._observe_text_start(event)
+        elif isinstance(event, TextMessageContentEvent):
+            self._observe_text_content(event)
+        elif isinstance(event, TextMessageEndEvent):
+            self._observe_text_end(event)
+        elif isinstance(event, ToolCallStartEvent):
+            self._observe_tool_call_start(event)
+        elif isinstance(event, ToolCallArgsEvent):
+            self._observe_tool_call_args(event)
+        elif isinstance(event, ToolCallResultEvent):
+            self._observe_tool_call_result(event)
+
+    def build(self) -> AGUIThreadSnapshot:
+        """Return the replayable thread snapshot."""
+        self._flush_open_text_message()
+        messages = self._emitted_messages if self._emitted_messages is not None else self._synthesized_messages
+        return AGUIThreadSnapshot(messages=messages, state=self.state, interrupt=self.interrupt)
+
+    def _observe_text_start(self, event: TextMessageStartEvent) -> None:
+        if self._open_text_message is not None and self._open_text_message.get("id") != event.message_id:
+            self._flush_open_text_message()
+        self._open_text_message = {"id": event.message_id, "role": event.role, "content": ""}
+
+    def _observe_text_content(self, event: TextMessageContentEvent) -> None:
+        if self._open_text_message is None or self._open_text_message.get("id") != event.message_id:
+            self._open_text_message = {"id": event.message_id, "role": "assistant", "content": ""}
+        self._open_text_message["content"] = f"{self._open_text_message.get('content', '')}{event.delta}"
+
+    def _observe_text_end(self, event: TextMessageEndEvent) -> None:
+        if self._open_text_message is None or self._open_text_message.get("id") != event.message_id:
+            return
+        self._flush_open_text_message()
+
+    def _observe_tool_call_start(self, event: ToolCallStartEvent) -> None:
+        parent_message_id = event.parent_message_id
+        if (
+            self._open_text_message is not None
+            and parent_message_id is not None
+            and self._open_text_message.get("id") == parent_message_id
+            and self._open_text_message.get("content")
+        ):
+            self._open_text_message["id"] = generate_event_id()
+        self._flush_open_text_message()
+        if self._tool_call_message is None or (
+            parent_message_id is not None and self._tool_call_message.get("id") != parent_message_id
+        ):
+            self._tool_call_message = {
+                "id": parent_message_id or generate_event_id(),
+                "role": "assistant",
+                "tool_calls": [],
+            }
+            self._synthesized_messages.append(self._tool_call_message)
+
+        tool_call = {
+            "id": event.tool_call_id,
+            "type": "function",
+            "function": {"name": event.tool_call_name, "arguments": ""},
+        }
+        cast(list[dict[str, Any]], self._tool_call_message["tool_calls"]).append(tool_call)
+        self._tool_calls_by_id[event.tool_call_id] = tool_call
+
+    def _observe_tool_call_args(self, event: ToolCallArgsEvent) -> None:
+        tool_call = self._tool_calls_by_id.get(event.tool_call_id)
+        if tool_call is None:
+            return
+        function_payload = cast(dict[str, Any], tool_call["function"])
+        function_payload["arguments"] = f"{function_payload.get('arguments', '')}{event.delta}"
+
+    def _observe_tool_call_result(self, event: ToolCallResultEvent) -> None:
+        self._synthesized_messages.append(
+            {
+                "id": event.message_id,
+                "role": "tool",
+                "toolCallId": event.tool_call_id,
+                "content": event.content,
+            }
+        )
+        # A result closes the current tool-call group; later tool calls start a new
+        # assistant message so replayed transcripts keep results adjacent to their
+        # tool_calls message, which provider APIs require.
+        self._tool_call_message = None
+
+    def _flush_open_text_message(self) -> None:
+        if self._open_text_message is None:
+            return
+        if self._open_text_message.get("content"):
+            self._synthesized_messages.append(self._open_text_message)
+            # Text between tool calls closes the current tool-call group as well.
+            self._tool_call_message = None
+        self._open_text_message = None
+
+
+async def _hydrate_workflow_thread_snapshot(
+    *,
+    snapshot_store: AGUIThreadSnapshotStore,
+    scope: str,
+    thread_id: str,
+    run_id: str,
+) -> AsyncGenerator[BaseEvent]:
+    """Replay the latest stored workflow AG-UI Thread Snapshot without invoking the workflow."""
+    yield RunStartedEvent(run_id=run_id, thread_id=thread_id)
+    snapshot = await snapshot_store.get(scope=scope, thread_id=thread_id)
+    if snapshot is None:
+        yield _build_run_finished_event(run_id=run_id, thread_id=thread_id)
+        return
+
+    if snapshot.state is not None:
+        yield StateSnapshotEvent(snapshot=snapshot.state)
+    if snapshot.messages:
+        yield MessagesSnapshotEvent(messages=snapshot.messages)  # type: ignore[arg-type]
+    yield _build_run_finished_event(run_id=run_id, thread_id=thread_id, interrupts=snapshot.interrupt)
+
+
 class AgentFrameworkWorkflow:
     """Base AG-UI workflow wrapper.
 
@@ -29,15 +214,30 @@ class AgentFrameworkWorkflow:
         workflow_factory: WorkflowFactory | None = None,
         name: str | None = None,
         description: str | None = None,
+        snapshot_store: AGUIThreadSnapshotStore | None = None,
     ) -> None:
+        """Initialize the AG-UI workflow wrapper.
+
+        Args:
+            workflow: Optional workflow instance to expose.
+            workflow_factory: Optional factory for thread-scoped workflow instances.
+            name: Optional workflow name.
+            description: Optional workflow description.
+            snapshot_store: Optional AG-UI Thread Snapshot store. Snapshot persistence remains inactive unless
+                endpoint setup also provides an explicit Snapshot Scope resolver.
+        """
         if workflow is not None and workflow_factory is not None:
             raise ValueError("Pass either workflow= or workflow_factory=, not both.")
 
         self.workflow = workflow
         self._workflow_factory = workflow_factory
-        self._workflow_by_thread: dict[str, Workflow] = {}
+        # Cache keyed by (snapshot_scope, thread_id): the Snapshot Scope is the
+        # authorization boundary, so the same thread id under different scopes
+        # must never share an in-memory workflow instance.
+        self._workflow_by_thread: dict[tuple[str | None, str], Workflow] = {}
         self.name = name if name is not None else getattr(workflow, "name", "workflow")
         self.description = description if description is not None else getattr(workflow, "description", "")
+        self.snapshot_store = snapshot_store
 
     @staticmethod
     def _thread_id_from_input(input_data: dict[str, Any]) -> str:
@@ -47,7 +247,7 @@ class AgentFrameworkWorkflow:
             return str(thread_id)
         return str(uuid.uuid4())
 
-    def _resolve_workflow(self, thread_id: str) -> Workflow:
+    def _resolve_workflow(self, thread_id: str, snapshot_scope: str | None = None) -> Workflow:
         """Get the workflow instance for the current run."""
         if self.workflow is not None:
             return self.workflow
@@ -55,17 +255,22 @@ class AgentFrameworkWorkflow:
         if self._workflow_factory is None:
             raise NotImplementedError("No workflow is attached. Override run or pass workflow=/workflow_factory=.")
 
-        workflow = self._workflow_by_thread.get(thread_id)
+        cache_key = (snapshot_scope, thread_id)
+        workflow = self._workflow_by_thread.get(cache_key)
         if workflow is None:
             workflow = self._workflow_factory(thread_id)
             if not isinstance(workflow, Workflow):
                 raise TypeError("workflow_factory must return a Workflow instance.")
-            self._workflow_by_thread[thread_id] = workflow
+            self._workflow_by_thread[cache_key] = workflow
         return workflow
 
-    def clear_thread_workflow(self, thread_id: str) -> None:
-        """Drop a single cached thread workflow instance."""
-        self._workflow_by_thread.pop(thread_id, None)
+    def clear_thread_workflow(self, thread_id: str, snapshot_scope: str | None = None) -> None:
+        """Drop cached workflow instances for a thread, optionally limited to one Snapshot Scope."""
+        if snapshot_scope is not None:
+            self._workflow_by_thread.pop((snapshot_scope, thread_id), None)
+            return
+        for key in [key for key in self._workflow_by_thread if key[1] == thread_id]:
+            del self._workflow_by_thread[key]
 
     def clear_workflow_cache(self) -> None:
         """Drop all cached thread workflow instances."""
@@ -77,6 +282,96 @@ class AgentFrameworkWorkflow:
         Subclasses may override this to provide custom AG-UI streams.
         """
         thread_id = self._thread_id_from_input(input_data)
-        workflow = self._resolve_workflow(thread_id)
+        run_id = str(input_data.get("run_id") or input_data.get("runId") or uuid.uuid4())
+        snapshot_scope = cast(str | None, input_data.get(_SNAPSHOT_SCOPE_INPUT_KEY))
+        raw_messages = list(cast(list[dict[str, Any]], input_data.get("messages", []) or []))
+        resume_payload = _extract_resume_payload(input_data)
+        snapshot_store = self.snapshot_store
+
+        if snapshot_store is not None and snapshot_scope is not None and not raw_messages and resume_payload is None:
+            async for event in _hydrate_workflow_thread_snapshot(
+                snapshot_store=snapshot_store,
+                scope=snapshot_scope,
+                thread_id=thread_id,
+                run_id=run_id,
+            ):
+                yield event
+            return
+
+        # Load the stored snapshot for follow-up turns so the workflow runs with the
+        # full persisted thread history instead of just the latest request messages.
+        stored_snapshot: AGUIThreadSnapshot | None = None
+        if snapshot_store is not None and snapshot_scope is not None:
+            stored_snapshot = await snapshot_store.get(scope=snapshot_scope, thread_id=thread_id)
+            if stored_snapshot is not None and resume_payload is None:
+                raw_messages = _reconstruct_messages_from_thread_snapshot(
+                    stored_messages=stored_snapshot.messages,
+                    incoming_messages=raw_messages,
+                    stored_interrupt=stored_snapshot.interrupt,
+                )
+                input_data["messages"] = raw_messages
+
+        # Merge stored state with request overrides, then fill endpoint-deferred
+        # defaults only for keys missing from both.
+        request_state = input_data.get("state")
+        deferred_default_state = cast(dict[str, Any] | None, input_data.get(_DEFAULT_STATE_INPUT_KEY))
+        effective_state: dict[str, Any] = {}
+        if stored_snapshot is not None and stored_snapshot.state is not None:
+            effective_state.update(stored_snapshot.state)
+        if isinstance(request_state, dict):
+            effective_state.update(cast(dict[str, Any], request_state))
+        if deferred_default_state:
+            for key, value in deferred_default_state.items():
+                if key not in effective_state:
+                    effective_state[key] = copy.deepcopy(value)
+        if effective_state:
+            input_data["state"] = effective_state
+
+        workflow = self._resolve_workflow(thread_id, snapshot_scope)
+        builder_seed_messages = raw_messages
+        if resume_payload is not None and stored_snapshot is not None:
+            # Resume requests carry only the synthesized interrupt response, so seed
+            # the builder with stored history to avoid persisting a truncated thread.
+            builder_seed_messages = [
+                copy.deepcopy(message) for message in stored_snapshot.messages
+            ] + builder_seed_messages
+        snapshot_builder = (
+            _WorkflowSnapshotBuilder(builder_seed_messages)
+            if snapshot_store is not None and snapshot_scope is not None
+            else None
+        )
+        if snapshot_builder is not None and effective_state:
+            # Seed builder state so a run that emits no StateSnapshotEvent still
+            # persists the latest known Shared State instead of dropping it.
+            state_snapshot = make_json_safe(effective_state)
+            if isinstance(state_snapshot, dict):
+                snapshot_builder.state = cast(dict[str, Any], state_snapshot)
+        run_error_emitted = False
         async for event in run_workflow_stream(input_data, workflow):
+            if snapshot_builder is not None:
+                snapshot_builder.observe(event)
+            if isinstance(event, RunErrorEvent):
+                run_error_emitted = True
             yield event
+
+        if (
+            snapshot_builder is not None
+            and not run_error_emitted
+            and snapshot_store is not None
+            and snapshot_scope is not None
+        ):
+            try:
+                await snapshot_store.save(
+                    scope=snapshot_scope,
+                    thread_id=thread_id,
+                    snapshot=snapshot_builder.build(),
+                )
+            except Exception:
+                # RUN_FINISHED has already been yielded; a store failure must not
+                # surface as a second terminal RUN_ERROR event. The previous
+                # snapshot stays available for hydration.
+                logger.exception(
+                    "Failed to save AG-UI Thread Snapshot for scope=%s thread_id=%s; keeping previous snapshot.",
+                    snapshot_scope,
+                    thread_id,
+                )

python/packages/ag-ui/tests/ag_ui/test_public_exports.py

@@ -32,6 +32,21 @@ def test_agent_framework_ag_ui_exports_state_update() -> None:
     assert callable(state_update)
 
 
+def test_agent_framework_ag_ui_exports_snapshot_primitives() -> None:
+    """Runtime package should export AG-UI Thread Snapshot primitives."""
+    from agent_framework_ag_ui import (
+        DEFAULT_MAX_THREAD_SNAPSHOTS,
+        AGUIThreadSnapshot,
+        AGUIThreadSnapshotStore,
+        InMemoryAGUIThreadSnapshotStore,
+    )
+
+    assert AGUIThreadSnapshot.__name__ == "AGUIThreadSnapshot"
+    assert AGUIThreadSnapshotStore.__name__ == "AGUIThreadSnapshotStore"
+    assert InMemoryAGUIThreadSnapshotStore.__name__ == "InMemoryAGUIThreadSnapshotStore"
+    assert DEFAULT_MAX_THREAD_SNAPSHOTS >= 1
+
+
 def test_core_ag_ui_lazy_exports_include_event_converter_and_http_service() -> None:
     """Core facade must expose AGUIEventConverter, AGUIHttpService, and __version__."""
     from agent_framework import ag_ui
@@ -39,3 +54,13 @@ def test_core_ag_ui_lazy_exports_include_event_converter_and_http_service() -> N
     assert hasattr(ag_ui, "AGUIEventConverter")
     assert hasattr(ag_ui, "AGUIHttpService")
     assert hasattr(ag_ui, "__version__")
+
+
+def test_core_ag_ui_lazy_exports_include_snapshot_primitives() -> None:
+    """Core facade must expose snapshot primitives needed for endpoint configuration."""
+    from agent_framework import ag_ui
+
+    assert hasattr(ag_ui, "AGUIThreadSnapshot")
+    assert hasattr(ag_ui, "AGUIThreadSnapshotStore")
+    assert hasattr(ag_ui, "InMemoryAGUIThreadSnapshotStore")
+    assert hasattr(ag_ui, "SnapshotScopeResolver")

python/packages/ag-ui/tests/ag_ui/test_snapshots.py

@@ -0,0 +1,160 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Tests for AG-UI thread snapshot storage primitives."""
+
+from dataclasses import fields
+
+from agent_framework_ag_ui import AGUIThreadSnapshot, AGUIThreadSnapshotStore, InMemoryAGUIThreadSnapshotStore
+
+
+def test_thread_snapshot_model_contains_only_replayable_snapshot_fields() -> None:
+    """The public snapshot model is limited to messages, Shared State, and interruption state."""
+    assert [field.name for field in fields(AGUIThreadSnapshot)] == ["messages", "state", "interrupt"]
+
+
+def test_in_memory_snapshot_store_satisfies_snapshot_store_protocol() -> None:
+    """The built-in store conforms to the public async store protocol."""
+    assert isinstance(InMemoryAGUIThreadSnapshotStore(), AGUIThreadSnapshotStore)
+
+
+async def test_in_memory_snapshot_store_replaces_latest_snapshot() -> None:
+    """Saving the same scoped thread key replaces the previous snapshot."""
+    store = InMemoryAGUIThreadSnapshotStore()
+
+    await store.save(
+        scope="tenant-a",
+        thread_id="thread-1",
+        snapshot=AGUIThreadSnapshot(messages=[{"id": "first"}], state={"count": 1}),
+    )
+    await store.save(
+        scope="tenant-a",
+        thread_id="thread-1",
+        snapshot=AGUIThreadSnapshot(messages=[{"id": "second"}], state={"count": 2}),
+    )
+
+    snapshot = await store.get(scope="tenant-a", thread_id="thread-1")
+
+    assert snapshot is not None
+    assert snapshot.messages == [{"id": "second"}]
+    assert snapshot.state == {"count": 2}
+
+
+async def test_in_memory_snapshot_store_keeps_scopes_separate() -> None:
+    """The same AG-UI Thread id in different Snapshot Scopes addresses different snapshots."""
+    store = InMemoryAGUIThreadSnapshotStore()
+
+    await store.save(
+        scope="tenant-a",
+        thread_id="thread-1",
+        snapshot=AGUIThreadSnapshot(messages=[{"id": "a", "role": "user", "content": "from a"}]),
+    )
+    await store.save(
+        scope="tenant-b",
+        thread_id="thread-1",
+        snapshot=AGUIThreadSnapshot(messages=[{"id": "b", "role": "user", "content": "from b"}]),
+    )
+
+    tenant_a_snapshot = await store.get(scope="tenant-a", thread_id="thread-1")
+    tenant_b_snapshot = await store.get(scope="tenant-b", thread_id="thread-1")
+
+    assert tenant_a_snapshot is not None
+    assert tenant_b_snapshot is not None
+    assert tenant_a_snapshot.messages == [{"id": "a", "role": "user", "content": "from a"}]
+    assert tenant_b_snapshot.messages == [{"id": "b", "role": "user", "content": "from b"}]
+
+
+async def test_in_memory_snapshot_store_deletes_and_clears_snapshots() -> None:
+    """Delete removes one scoped thread key, while clear can remove a scope or the whole store."""
+    store = InMemoryAGUIThreadSnapshotStore()
+
+    await store.save(scope="tenant-a", thread_id="thread-1", snapshot=AGUIThreadSnapshot(messages=[{"id": "a1"}]))
+    await store.save(scope="tenant-a", thread_id="thread-2", snapshot=AGUIThreadSnapshot(messages=[{"id": "a2"}]))
+    await store.save(scope="tenant-b", thread_id="thread-1", snapshot=AGUIThreadSnapshot(messages=[{"id": "b1"}]))
+
+    assert await store.delete(scope="tenant-a", thread_id="thread-1") is True
+    assert await store.delete(scope="tenant-a", thread_id="thread-1") is False
+    assert await store.get(scope="tenant-a", thread_id="thread-1") is None
+    assert await store.get(scope="tenant-a", thread_id="thread-2") is not None
+
+    await store.clear(scope="tenant-a")
+
+    assert await store.get(scope="tenant-a", thread_id="thread-2") is None
+    assert await store.get(scope="tenant-b", thread_id="thread-1") is not None
+
+    await store.clear()
+
+    assert await store.get(scope="tenant-b", thread_id="thread-1") is None
+
+
+async def test_in_memory_snapshot_store_evicts_oldest_snapshot_when_bounded() -> None:
+    """The memory store bounds retained scoped thread snapshots."""
+    store = InMemoryAGUIThreadSnapshotStore(max_snapshots=2)
+
+    await store.save(scope="tenant-a", thread_id="thread-1", snapshot=AGUIThreadSnapshot(messages=[{"id": "first"}]))
+    await store.save(scope="tenant-a", thread_id="thread-2", snapshot=AGUIThreadSnapshot(messages=[{"id": "second"}]))
+    await store.save(scope="tenant-a", thread_id="thread-3", snapshot=AGUIThreadSnapshot(messages=[{"id": "third"}]))
+
+    assert await store.get(scope="tenant-a", thread_id="thread-1") is None
+    assert await store.get(scope="tenant-a", thread_id="thread-2") is not None
+    assert await store.get(scope="tenant-a", thread_id="thread-3") is not None
+
+
+def test_workflow_snapshot_builder_splits_tool_call_groups() -> None:
+    """Tool calls separated by results or text synthesize provider-valid message groups."""
+    from ag_ui.core import (
+        TextMessageContentEvent,
+        TextMessageEndEvent,
+        TextMessageStartEvent,
+        ToolCallArgsEvent,
+        ToolCallResultEvent,
+        ToolCallStartEvent,
+    )
+
+    from agent_framework_ag_ui._workflow import _WorkflowSnapshotBuilder
+
+    builder = _WorkflowSnapshotBuilder([])
+    builder.observe(ToolCallStartEvent(tool_call_id="call-a", tool_call_name="toolA"))
+    builder.observe(ToolCallArgsEvent(tool_call_id="call-a", delta='{"x": 1}'))
+    builder.observe(ToolCallResultEvent(message_id="result-a", tool_call_id="call-a", content="resA"))
+    builder.observe(TextMessageStartEvent(message_id="text-1", role="assistant"))
+    builder.observe(TextMessageContentEvent(message_id="text-1", delta="thinking"))
+    builder.observe(TextMessageEndEvent(message_id="text-1"))
+    builder.observe(ToolCallStartEvent(tool_call_id="call-b", tool_call_name="toolB"))
+    builder.observe(ToolCallResultEvent(message_id="result-b", tool_call_id="call-b", content="resB"))
+
+    messages = builder.build().messages
+    shapes = [
+        (
+            message.get("role"),
+            [tool_call["id"] for tool_call in message.get("tool_calls", [])] or message.get("toolCallId"),
+        )
+        for message in messages
+    ]
+    assert shapes == [
+        ("assistant", ["call-a"]),
+        ("tool", "call-a"),
+        ("assistant", None),
+        ("assistant", ["call-b"]),
+        ("tool", "call-b"),
+    ]
+
+
+async def test_in_memory_snapshot_store_rejects_invalid_keys() -> None:
+    """Key parts must be non-empty strings for every store operation."""
+    import pytest
+
+    store = InMemoryAGUIThreadSnapshotStore()
+    snapshot = AGUIThreadSnapshot()
+
+    with pytest.raises(ValueError):
+        await store.save(scope="", thread_id="thread-1", snapshot=snapshot)
+    with pytest.raises(ValueError):
+        await store.save(scope="tenant-a", thread_id="", snapshot=snapshot)
+    with pytest.raises(TypeError):
+        await store.save(scope=123, thread_id="thread-1", snapshot=snapshot)  # type: ignore[arg-type]
+    with pytest.raises(ValueError):
+        await store.get(scope="tenant-a", thread_id="")
+    with pytest.raises(TypeError):
+        await store.delete(scope=None, thread_id="thread-1")  # type: ignore[arg-type]
+    with pytest.raises(ValueError):
+        await store.clear(scope="")

python/packages/anthropic/agent_framework_anthropic/_chat_client.py

@@ -1024,8 +1024,10 @@ class RawAnthropicClient(
             usage_details["input_token_count"] = usage.input_tokens
         if usage.cache_creation_input_tokens is not None:
             usage_details["anthropic.cache_creation_input_tokens"] = usage.cache_creation_input_tokens  # type: ignore[typeddict-unknown-key]
+            usage_details["cache_creation_input_token_count"] = usage.cache_creation_input_tokens
         if usage.cache_read_input_tokens is not None:
             usage_details["anthropic.cache_read_input_tokens"] = usage.cache_read_input_tokens  # type: ignore[typeddict-unknown-key]
+            usage_details["cache_read_input_token_count"] = usage.cache_read_input_tokens
         return usage_details
 
     def _parse_contents_from_anthropic(

python/packages/anthropic/tests/test_anthropic_client.py

@@ -2354,6 +2354,27 @@ def test_parse_usage_with_cache_tokens(mock_anthropic_client: MagicMock) -> None
     assert result["input_token_count"] == 100
     assert result["anthropic.cache_creation_input_tokens"] == 20
     assert result["anthropic.cache_read_input_tokens"] == 30
+    assert result["cache_creation_input_token_count"] == 20
+    assert result["cache_read_input_token_count"] == 30
+
+
+def test_parse_usage_preserves_zero_cache_tokens(mock_anthropic_client: MagicMock) -> None:
+    """Test parsing usage preserves zero-valued mapped cache tokens."""
+    client = create_test_anthropic_client(mock_anthropic_client)
+
+    mock_usage = MagicMock()
+    mock_usage.input_tokens = 100
+    mock_usage.output_tokens = 50
+    mock_usage.cache_creation_input_tokens = 0
+    mock_usage.cache_read_input_tokens = 0
+
+    result = client._parse_usage_from_anthropic(mock_usage)
+
+    assert result is not None
+    assert result["anthropic.cache_creation_input_tokens"] == 0
+    assert result["cache_creation_input_token_count"] == 0
+    assert result["anthropic.cache_read_input_tokens"] == 0
+    assert result["cache_read_input_token_count"] == 0
 
 
 # Code Execution Result Tests

python/packages/core/AGENTS.md

@@ -94,11 +94,11 @@ agent_framework/
 
 ### File Access Harness (`_harness/_file_access.py`)
 
-- **`AgentFileStore`** - Abstract async store backing the file-access harness. Implementations expose `write_file`, `read_file`, `delete_file`, `list_files`, `file_exists`, `search_files`, and `create_directory` over forward-slash relative paths.
+- **`AgentFileStore`** - Abstract async store backing the file-access harness. Implementations expose `write_file`, `read_file`, `delete_file`, `list_files`, `list_directories`, `file_exists`, `search_files`, and `create_directory` over forward-slash relative paths. `list_files`/`list_directories` return only direct children; `search_files` accepts a keyword-only `recursive` flag (default `False`) and, when `recursive=True`, walks all descendants and returns `file_name` values relative to the search directory.
 - **`InMemoryAgentFileStore`** - Dict-backed store suitable for tests and lightweight scenarios.
 - **`FileSystemAgentFileStore`** - Disk-backed store rooted under a configurable directory. Enforces relative-path normalization, root containment, and rejects symlink/reparse-point segments to prevent escape.
 - **`FileSearchResult`** / **`FileSearchMatch`** - `SerializationMixin` DTOs returned by `search_files`, carrying the matching file name, a context snippet, and the matching lines with 1-based line numbers.
-- **`FileAccessProvider`** - `ContextProvider` that adds shared file-access tools (`file_access_save_file`, `file_access_read_file`, `file_access_delete_file`, `file_access_list_files`, `file_access_search_files`) plus default usage instructions to each invocation. Unlike `MemoryContextProvider`, the store is intentionally shared across sessions and agents.
+- **`FileAccessProvider`** - `ContextProvider` that adds shared file-access tools (`file_access_save_file`, `file_access_read_file`, `file_access_delete_file`, `file_access_list_files`, `file_access_list_subdirectories`, `file_access_search_files`) plus default usage instructions to each invocation. `file_access_list_files`/`file_access_list_subdirectories` enumerate direct children (files / subdirectories) so the agent can walk the tree level by level; `file_access_search_files` searches recursively from the store root and returns store-root-relative `file_name` paths, scoped via an `fnmatch` glob (where `*` crosses `/`, e.g. `*.md`, `reports/*`). Unlike `MemoryContextProvider`, the store is intentionally shared across sessions and agents.
 
 ### Tool Approval Harness (`_harness/_tool_approval.py`)
 
@@ -116,6 +116,11 @@ agent_framework/
   available, approval requests for known non-approval-required tools are treated as already approved, hidden, stored
   in session state keyed to the visible approval request ids from that batch, and reinjected only when that visible
   approval flow resumes.
+### Agent Loop (`_harness/_loop.py`)
+
+- **`AgentLoopMiddleware`** - `AgentMiddleware` that re-runs an agent in a loop by calling `call_next()` repeatedly (the pipeline re-reads `context.messages` each time). One configurable class covers two patterns: a required user `should_continue` predicate (sync or async, the first positional/keyword arg), and a chat-client judge built via the `.with_judge(...)` factory (a second chat client decides whether the original request was answered; loops while it is *not*, using a `JudgeVerdict` structured-output response — internally just an async `should_continue` predicate). The constructor covers the predicate pattern directly; only the judge has a convenience classmethod factory (`.with_judge(judge_client, ...)`) that forwards to `__init__`. Supports both streaming and non-streaming runs. By default a non-streaming run returns an aggregated `AgentResponse` containing every iteration's messages plus the injected `next_message` "nudge" messages (as `user` messages); set `return_final_only=True` to return only the last iteration's response. Streaming runs always yield each iteration's updates and emit the injected nudge messages as `user` updates between iterations (the `return_final_only` flag has no effect on streaming, and the final response reflects the last iteration; `MiddlewareTermination` is handled cleanly). `should_continue` is required; other constructor args are optional: `max_iterations` (safety cap; defaults to `DEFAULT_MAX_ITERATIONS`=10, explicit `None`→unbounded, positive int caps; `.with_judge` uses `DEFAULT_JUDGE_MAX_ITERATIONS`=5 as its default), `next_message` (defaults to a short "continue" nudge), `return_final_only`, and `additional_instructions` (an extra `system` message injected ahead of the input before the agent runs — becomes part of the original messages so it survives `fresh_context` resets and persists via a session). The judge is configured only through `.with_judge` (`judge_client`/`instructions`/`criteria`), not the constructor, and its `reasoning` is fed back to the agent as the next iteration's input; the judge forwards the original request messages and the agent's latest response messages verbatim so multi-modal content is preserved. `criteria` (a `list[str]`) is both injected as the agent's `additional_instructions` and rendered into the judge instructions wherever the `{{criteria}}` placeholder (`CRITERIA_PLACEHOLDER`) appears (`DEFAULT_JUDGE_INSTRUCTIONS` ends with it; custom `instructions` may include it, and it is stripped when no criteria are given). The `should_continue`/`next_message` callables are invoked with keyword args (`iteration`, `last_result`, `messages`, `original_messages`, `session`, `agent`, `progress`, `feedback`) and may be sync or async; declare only what you need plus `**kwargs`. `should_continue` may return a plain `bool` or a `(bool, str | None)` tuple whose second item is feedback surfaced to `next_message`/`record_feedback` via the `feedback` kwarg (the judge uses this to relay its `reasoning`). Stop precedence per iteration is `max_iterations` → `should_continue`, evaluated before `record_feedback` so the feedback is available to it.
+  - **Feedback tracking** - `record_feedback` captures a per-iteration progress entry (called with the loop kwargs; if it returns a truthy string the entry is appended, otherwise the agent's response text is used as the fallback entry). The accumulated log is exposed to every callback via the `progress` keyword (a per-iteration copy of prior entries) and, when `inject_progress=True` (default), injected into the next iteration's input as a `user` message (the full log without a session, only the latest entry with a session to avoid duplicating history). `fresh_context=True` restarts each iteration from the original task plus the progress log; when a session is attached it is snapshotted (`to_dict()`) before the loop and restored (`from_dict` + field copy) between iterations so the local transcript and any service-side conversation id reset too (in-loop working-state is discarded, pre-loop state preserved, continuity carried only by the progress log).
+- **`todos_remaining(provider)`** / **`background_tasks_running(provider)`** - Helper factories returning `should_continue` predicates that loop while a `TodoProvider` has open items, or while a `BackgroundAgentsProvider`'s persisted state shows running tasks.
 
 ### Workflows (`_workflows/`)
 

python/packages/core/agent_framework/__init__.py

@@ -102,6 +102,12 @@ from ._harness._file_access import (
     FileSystemAgentFileStore,
     InMemoryAgentFileStore,
 )
+from ._harness._loop import (
+    AgentLoopMiddleware,
+    JudgeVerdict,
+    background_tasks_running,
+    todos_remaining,
+)
 from ._harness._memory import (
     DEFAULT_MEMORY_SOURCE_ID,
     MemoryContextProvider,
@@ -363,6 +369,7 @@ __all__ = [
     "AgentExecutorResponse",
     "AgentFileStore",
     "AgentFrameworkException",
+    "AgentLoopMiddleware",
     "AgentMiddleware",
     "AgentMiddlewareLayer",
     "AgentMiddlewareTypes",
@@ -455,6 +462,7 @@ __all__ = [
     "InlineSkill",
     "InlineSkillResource",
     "InlineSkillScript",
+    "JudgeVerdict",
     "LocalEvaluator",
     "MCPSkill",
     "MCPSkillResource",
@@ -558,6 +566,7 @@ __all__ = [
     "agent_middleware",
     "annotate_message_groups",
     "apply_compaction",
+    "background_tasks_running",
     "chat_middleware",
     "create_always_approve_tool_response",
     "create_always_approve_tool_with_arguments_response",
@@ -588,6 +597,7 @@ __all__ = [
     "response_handler",
     "set_agent_mode",
     "step",
+    "todos_remaining",
     "tool",
     "tool_call_args_match",
     "tool_called_check",

python/packages/core/agent_framework/_harness/_file_access.py

@@ -5,9 +5,10 @@
 Unlike :class:`~agent_framework.MemoryContextProvider`, which provides
 session-scoped memory that may be isolated per session, :class:`FileAccessProvider`
 operates on a shared, persistent storage area whose contents are visible across
-sessions and agents. The provider exposes five tools — ``file_access_save_file``,
+sessions and agents. The provider exposes six tools — ``file_access_save_file``,
 ``file_access_read_file``, ``file_access_delete_file``, ``file_access_list_files``,
-and ``file_access_search_files`` — by registering them on the per-invocation
+``file_access_list_subdirectories``, and ``file_access_search_files`` — by
+registering them on the per-invocation
 :class:`~agent_framework.SessionContext` in :meth:`FileAccessProvider.before_run`.
 
 The store abstraction is generic so callers can plug in in-memory, local-disk, or
@@ -48,7 +49,11 @@ DEFAULT_FILE_ACCESS_INSTRUCTIONS = (
     "Use these tools to read input data provided by the user, write output "
     "artifacts, and manage any files the user has asked you to work with.\n\n"
     "- Never delete or overwrite existing files unless the user has explicitly "
-    "asked you to do so."
+    "asked you to do so.\n"
+    "- Files may be organized into subdirectories. Use `file_access_list_files` "
+    "and `file_access_list_subdirectories` to explore the tree level by level, "
+    "or `file_access_search_files` to search file contents recursively across "
+    "the whole store."
 )
 
 # Maximum number of characters of context to include on either side of the first
@@ -178,10 +183,16 @@ def _normalize_relative_path(path: str, *, is_directory: bool = False) -> str:
 def _matches_glob(file_name: str, pattern: str | None) -> bool:
     """Return whether ``file_name`` matches the optional glob pattern (case-insensitive).
 
-    When ``pattern`` is ``None`` or blank this returns True so callers can skip
-    filtering by passing nothing. Matching uses :func:`fnmatch.fnmatchcase` over a
-    lowercased pattern/name pair to give consistent results across operating
-    systems (``fnmatch.fnmatch`` is case-sensitive on POSIX but not on Windows).
+    ``file_name`` is the forward-slash path of a file relative to the search
+    directory (for a direct child this is just its basename; for a recursive
+    search it may contain ``/`` separators). When ``pattern`` is ``None`` or blank
+    this returns True so callers can skip filtering by passing nothing. Matching
+    uses :func:`fnmatch.fnmatchcase` over a lowercased pattern/name pair to give
+    consistent results across operating systems (``fnmatch.fnmatch`` is
+    case-sensitive on POSIX but not on Windows). Note that with ``fnmatch`` a
+    ``*`` matches any characters **including** ``/``, so ``"*.md"`` matches
+    markdown files at any depth and ``"reports/*"`` matches everything under
+    ``reports``.
     """
     if pattern is None or not pattern.strip():
         return True
@@ -418,6 +429,18 @@ class AgentFileStore(ABC):
             The list of file names (not full paths) in the specified directory.
         """
 
+    @abstractmethod
+    async def list_directories(self, directory: str = "") -> list[str]:
+        """List the direct child subdirectory names of ``directory``.
+
+        Args:
+            directory: The relative directory path to list. Use ``""`` for the root.
+
+        Returns:
+            The list of subdirectory names (not full paths) directly contained in
+            the specified directory.
+        """
+
     @abstractmethod
     async def file_exists(self, path: str) -> bool:
         """Return whether a file exists at ``path``.
@@ -432,6 +455,8 @@ class AgentFileStore(ABC):
         directory: str,
         regex_pattern: str,
         file_pattern: str | None = None,
+        *,
+        recursive: bool = False,
     ) -> list[FileSearchResult]:
         """Search files in ``directory`` for content matching ``regex_pattern``.
 
@@ -441,12 +466,19 @@ class AgentFileStore(ABC):
                 (case-insensitive). For example, ``"error|warning"`` matches lines
                 containing ``"error"`` or ``"warning"``.
             file_pattern: An optional glob pattern (case-insensitive) used to
-                filter which files are searched. When ``None`` or blank, every
-                file in the directory is searched.
+                filter which files are searched. The pattern is matched against
+                each file's path **relative to** ``directory`` (forward slashes).
+                When ``None`` or blank, every file in scope is searched.
+
+        Keyword Args:
+            recursive: When ``False`` (default) only the direct children of
+                ``directory`` are searched. When ``True`` every descendant file is
+                searched.
 
         Returns:
             The list of files whose content matched, with snippet and matching
-            line metadata.
+            line metadata. Each result's ``file_name`` is the path relative to
+            ``directory`` (forward slashes).
         """
 
     @abstractmethod
@@ -530,6 +562,38 @@ class InMemoryAgentFileStore(AgentFileStore):
             results.append(display[len(prefix) :])
         return results
 
+    async def list_directories(self, directory: str = "") -> list[str]:
+        """Return the direct child subdirectory names of ``directory``.
+
+        A subdirectory is the first path segment of any stored key whose
+        remainder (after the directory prefix) still contains a ``/`` separator.
+        Distinct first segments are collected, preserving the *original-case*
+        display name and de-duplicating case-insensitively, mirroring the
+        case-preserving behaviour of :class:`FileSystemAgentFileStore`.
+        """
+        prefix = _normalize_relative_path(directory, is_directory=True).lower()
+        if prefix and not prefix.endswith("/"):
+            prefix += "/"
+        async with self._lock:
+            entries = [(key, display) for key, (display, _) in self._files.items()]
+        results: list[str] = []
+        seen: set[str] = set()
+        for key, display in entries:
+            if not key.startswith(prefix):
+                continue
+            remainder = key[len(prefix) :]
+            separator_index = remainder.find("/")
+            if separator_index <= 0:
+                continue
+            segment_key = remainder[:separator_index]
+            if segment_key in seen:
+                continue
+            seen.add(segment_key)
+            # ``display`` is the original-case normalized path; take the matching
+            # first segment after the (case-insensitive) prefix.
+            results.append(display[len(prefix) : len(prefix) + separator_index])
+        return results
+
     async def file_exists(self, path: str) -> bool:
         """Return whether the file exists."""
         key = self._key(path)
@@ -541,6 +605,8 @@ class InMemoryAgentFileStore(AgentFileStore):
         directory: str,
         regex_pattern: str,
         file_pattern: str | None = None,
+        *,
+        recursive: bool = False,
     ) -> list[FileSearchResult]:
         """Search file contents for ``regex_pattern`` matches.
 
@@ -548,7 +614,10 @@ class InMemoryAgentFileStore(AgentFileStore):
         to a worker thread with a bounded timeout so a pathological pattern
         cannot stall the event loop. Returned :class:`FileSearchResult`
         instances use the *original-case* file names so the result mirrors
-        what :class:`FileSystemAgentFileStore` would produce.
+        what :class:`FileSystemAgentFileStore` would produce. The glob and each
+        result's ``file_name`` are relative to ``directory``; when ``recursive``
+        is ``True`` all descendants are searched and the relative path may
+        contain ``/`` separators.
         """
         prefix = _normalize_relative_path(directory, is_directory=True).lower()
         if prefix and not prefix.endswith("/"):
@@ -564,7 +633,7 @@ class InMemoryAgentFileStore(AgentFileStore):
                 if not key.startswith(prefix):
                     continue
                 relative_key = key[len(prefix) :]
-                if "/" in relative_key:
+                if not recursive and "/" in relative_key:
                     continue
                 relative_display = display[len(prefix) :]
                 if not _matches_glob(relative_display, file_pattern):
@@ -795,6 +864,28 @@ class FileSystemAgentFileStore(AgentFileStore):
                 names.append(entry.name)
         return names
 
+    async def list_directories(self, directory: str = "") -> list[str]:
+        """Return the direct child subdirectory names of ``directory``.
+
+        Symlinked directories (and reparse points on Windows) are excluded so a
+        listing cannot surface a path that escapes the root. An empty list is
+        returned for a non-existent directory.
+        """
+        full_dir = self._resolve_safe_directory_path(directory)
+        return await asyncio.to_thread(self._list_directories_sync, full_dir)
+
+    @staticmethod
+    def _list_directories_sync(full_dir: Path) -> list[str]:
+        if not full_dir.is_dir():
+            return []
+        names: list[str] = []
+        for entry in full_dir.iterdir():
+            if entry.is_symlink():
+                continue
+            if entry.is_dir():
+                names.append(entry.name)
+        return names
+
     async def file_exists(self, path: str) -> bool:
         """Return whether the file exists."""
         full_path = self._resolve_safe_path(path)
@@ -809,6 +900,8 @@ class FileSystemAgentFileStore(AgentFileStore):
         directory: str,
         regex_pattern: str,
         file_pattern: str | None = None,
+        *,
+        recursive: bool = False,
     ) -> list[FileSearchResult]:
         """Search file contents for ``regex_pattern`` matches.
 
@@ -816,23 +909,50 @@ class FileSystemAgentFileStore(AgentFileStore):
         file does not abort the whole directory search). Each skip is logged at
         ``WARNING`` level and a summary is logged at ``INFO`` so operators can
         tell the difference between "no matches" and "the corpus was largely
-        not searchable".
+        not searchable". The glob and each result's ``file_name`` are the file's
+        path relative to ``directory`` (forward slashes); when ``recursive`` is
+        ``True`` all descendant files are searched, otherwise only the direct
+        children.
         """
         full_dir = self._resolve_safe_directory_path(directory)
         regex = _compile_search_regex(regex_pattern)
-        return await _run_search_with_timeout(lambda: self._search_files_sync(full_dir, regex, file_pattern))
+        return await _run_search_with_timeout(lambda: self._search_files_sync(full_dir, regex, file_pattern, recursive))
 
     @staticmethod
-    def _search_files_sync(full_dir: Path, regex: re.Pattern[str], file_pattern: str | None) -> list[FileSearchResult]:
+    def _enumerate_search_files(full_dir: Path, recursive: bool) -> list[tuple[str, Path]]:
+        """Enumerate ``(relative_name, path)`` for files to search under ``full_dir``.
+
+        Symlinked files and symlinked directories (reparse points on Windows)
+        are skipped so the search cannot read or descend outside the root.
+        ``relative_name`` is the file's path relative to ``full_dir`` using
+        forward slashes.
+        """
+        found: list[tuple[str, Path]] = []
+        directories: list[Path] = [full_dir]
+        while directories:
+            current = directories.pop()
+            for entry in current.iterdir():
+                if entry.is_symlink():
+                    continue
+                if entry.is_dir():
+                    if recursive:
+                        directories.append(entry)
+                    continue
+                if entry.is_file():
+                    relative_name = entry.relative_to(full_dir).as_posix()
+                    found.append((relative_name, entry))
+        return found
+
+    @staticmethod
+    def _search_files_sync(
+        full_dir: Path, regex: re.Pattern[str], file_pattern: str | None, recursive: bool
+    ) -> list[FileSearchResult]:
         if not full_dir.is_dir():
             return []
         results: list[FileSearchResult] = []
         skipped: list[str] = []
-        for entry in full_dir.iterdir():
-            if entry.is_symlink() or not entry.is_file():
-                continue
-            file_name = entry.name
-            if not _matches_glob(file_name, file_pattern):
+        for relative_name, entry in FileSystemAgentFileStore._enumerate_search_files(full_dir, recursive):
+            if not _matches_glob(relative_name, file_pattern):
                 continue
             try:
                 file_content = entry.read_text(encoding="utf-8")
@@ -841,9 +961,9 @@ class FileSystemAgentFileStore(AgentFileStore):
                 # un-decodable entry doesn't abort the whole directory search.
                 # Log per file so operators can audit which files were skipped.
                 logger.warning("Skipping non-UTF-8 file during search: %s", entry)
-                skipped.append(file_name)
+                skipped.append(relative_name)
                 continue
-            result = _search_file_content(file_name, file_content, regex)
+            result = _search_file_content(relative_name, file_content, regex)
             if result is not None:
                 results.append(result)
         if skipped:
@@ -865,15 +985,18 @@ class FileSystemAgentFileStore(AgentFileStore):
 class FileAccessProvider(ContextProvider):
     """Context provider that gives an agent CRUD/search access to a shared file store.
 
-    The provider exposes five tools to the agent via the per-invocation
+    The provider exposes six tools to the agent via the per-invocation
     :class:`~agent_framework.SessionContext`:
 
     - ``file_access_save_file`` — Save a file (refuses to overwrite by default).
     - ``file_access_read_file`` — Read the content of a file by name.
     - ``file_access_delete_file`` — Delete a file by name.
-    - ``file_access_list_files`` — List all file names at the store root.
-    - ``file_access_search_files`` — Search file contents using a case-insensitive
-      regex, optionally filtered by a glob pattern over file names.
+    - ``file_access_list_files`` — List the direct child file names of a directory.
+    - ``file_access_list_subdirectories`` — List the direct child subdirectory
+      names of a directory.
+    - ``file_access_search_files`` — Recursively search file contents from the
+      store root using a case-insensitive regex, optionally filtered by a glob
+      pattern over the store-root-relative file paths.
 
     Unlike :class:`~agent_framework.MemoryContextProvider`, which provides
     session-scoped memory that may be isolated per session,
@@ -976,17 +1099,45 @@ class FileAccessProvider(ContextProvider):
             except OSError as exc:
                 return f"Could not list directory '{directory or ''}': {exc.strerror or exc}"
 
+        @tool(name="file_access_list_subdirectories", approval_mode="never_require")
+        async def file_access_list_subdirectories(directory: str | None = None) -> list[str] | str:
+            """List the direct child subdirectory names of a directory.
+
+            Omit ``directory`` (or pass an empty string) to list the root.
+            To enumerate subdirectories of a subdirectory, pass its relative path, for example
+            ``"reports"`` or ``"reports/2024"``.
+            Use this together with file_access_list_files to explore the directory tree level by level.
+            """
+            target = directory if directory and directory.strip() else ""
+            try:
+                return await self.store.list_directories(target)
+            except ValueError as exc:
+                return f"Could not list directory '{directory or ''}': {exc}"
+            except OSError as exc:
+                return f"Could not list directory '{directory or ''}': {exc.strerror or exc}"
+
         @tool(name="file_access_search_files", approval_mode="never_require")
         async def file_access_search_files(
             regex_pattern: str,
             file_pattern: str | None = None,
-            directory: str | None = None,
         ) -> list[dict[str, Any]] | str:
-            """Search file contents using a regular expression pattern (case-insensitive). Optionally filter which files to search using a glob pattern (e.g., "*.md", "research*"). Optionally scope the search to a subdirectory by passing its relative path; omit ``directory`` (or pass an empty string) to search the root. Returns matching file names, snippets, and matching lines with line numbers. The regex_pattern must be 256 characters or fewer."""  # noqa: E501
+            """Search the contents of all files in the store using a case-insensitive regular expression.
+
+            The search runs recursively across all subdirectories.
+            Optionally filter which files to search using a glob pattern matched against each file's
+            path relative to the store root.
+            The glob uses fnmatch semantics where ``*`` matches any characters including ``/``: use
+            ``"*.md"`` to match markdown files at any depth,
+            or ``"reports/*"`` to restrict the search to the ``reports`` subtree.
+            Leave empty or omit to search all files.
+            Returns matching results whose file_name values are paths relative to the store root
+            (usable with file_access_read_file),
+            along with snippets and matching lines with line numbers. The regex_pattern must be
+            256 characters or fewer.
+            """
             pattern = file_pattern if file_pattern and file_pattern.strip() else None
-            target = directory if directory and directory.strip() else ""
             try:
-                results = await self.store.search_files(target, regex_pattern, pattern)
+                results = await self.store.search_files("", regex_pattern, pattern, recursive=True)
             except ValueError as exc:
                 return f"Could not search files: {exc}"
             except OSError as exc:
@@ -1001,6 +1152,7 @@ class FileAccessProvider(ContextProvider):
                 file_access_read_file,
                 file_access_delete_file,
                 file_access_list_files,
+                file_access_list_subdirectories,
                 file_access_search_files,
             ],
         )

python/packages/core/agent_framework/_harness/_loop.py

@@ -0,0 +1,796 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""AgentLoopMiddleware: re-run an agent in a loop until a criterion is met.
+
+This module provides :class:`AgentLoopMiddleware`, an :class:`~agent_framework.AgentMiddleware`
+that repeatedly re-invokes the wrapped agent while a ``should_continue`` predicate says to keep
+going. It serves two common patterns through a single configurable class:
+
+1. A user-supplied ``should_continue`` predicate - for example, keep looping while a response does
+   not yet contain a completion marker, while a :class:`~agent_framework.TodoProvider` still has
+   open items, or while a :class:`~agent_framework.BackgroundAgentsProvider` still has running
+   tasks (see the :func:`todos_remaining` and :func:`background_tasks_running` helpers). The loop
+   can track a **feedback log** across iterations (``record_feedback``): each pass contributes an
+   entry that is exposed to every callback via the ``progress`` keyword and (by default) injected
+   into the next iteration's input. Set ``fresh_context=True`` to restart each pass from the
+   original task plus the progress log (with a session attached, the session is also snapshotted
+   before the loop and restored between iterations so no accumulated history leaks back in).
+   ``max_iterations`` bounds the loop as a safety cap.
+2. A chat-client judge (via :meth:`AgentLoopMiddleware.with_judge`) - a second chat client decides
+   whether the user's original request has been answered (via a :class:`JudgeVerdict` structured
+   output); the loop continues while the answer is "no". This is a convenience wrapper that builds an
+   async ``should_continue`` predicate, so it is a special case of (1).
+
+In every case, the input for the next iteration is controlled by the ``next_message`` callable.
+"""
+
+from __future__ import annotations
+
+import inspect
+from collections.abc import Awaitable, Callable, Sequence
+from typing import TYPE_CHECKING, Any, TypeAlias
+
+from pydantic import BaseModel, Field
+from typing_extensions import Self
+
+from .._feature_stage import ExperimentalFeature, experimental
+from .._middleware import AgentContext, AgentMiddleware, MiddlewareTermination
+from .._types import (
+    AgentResponse,
+    AgentResponseUpdate,
+    AgentRunInputs,
+    Message,
+    ResponseStream,
+    UsageDetails,
+    add_usage_details,
+    normalize_messages,
+)
+
+if TYPE_CHECKING:
+    from .._clients import SupportsChatGetResponse
+
+__all__ = [
+    "AgentLoopMiddleware",
+    "JudgeVerdict",
+    "background_tasks_running",
+    "todos_remaining",
+]
+
+DEFAULT_NEXT_MESSAGE = "Continue working on the task. If it is complete, say so."
+
+# Placeholder substituted with the rendered ``criteria`` block in judge instructions (see
+# :meth:`AgentLoopMiddleware.with_judge`). User-supplied instructions may include it to control
+# where the criteria are inserted; if absent, the criteria are not added to the judge instructions.
+CRITERIA_PLACEHOLDER = "{{criteria}}"
+
+# Verdict markers the judge is asked to emit for clients that do not honor structured output. They
+# are deliberately non-overlapping: neither marker is a substring of the other, nor of the JSON
+# field name ``answered``, so the text fallback in :func:`_build_judge_condition` cannot misclassify
+# a negative verdict (e.g. ``{"answered": false}``) as a positive one.
+JUDGE_VERDICT_DONE = "VERDICT: DONE"
+JUDGE_VERDICT_MORE = "VERDICT: MORE"
+
+DEFAULT_JUDGE_INSTRUCTIONS = (
+    "You are an evaluator. You are given a user's original request and an agent's latest response. "
+    "Decide whether the agent has fully addressed the original request. "
+    "Set 'answered' to true if the request has been fully addressed, or false if more work is still "
+    "required, and use 'reasoning' to briefly justify your decision. "
+    f"If you cannot return structured output, end your reply with a line reading exactly "
+    f"'{JUDGE_VERDICT_DONE}' when the request has been fully addressed or '{JUDGE_VERDICT_MORE}' "
+    f"when more work is still required."
+    "{{criteria}}"
+)
+
+
+def _render_criteria_block(criteria: Sequence[str] | None) -> str:
+    """Render a list of criteria into a bullet block for the judge instructions (``""`` if none)."""
+    if not criteria:
+        return ""
+    bullets = "\n".join(f"- {item}" for item in criteria)
+    return f"\n\nThe response must satisfy all of the following criteria:\n{bullets}"
+
+
+def _criteria_agent_instruction(criteria: Sequence[str]) -> str:
+    """Render the criteria into an extra instruction injected for the agent before each run."""
+    bullets = "\n".join(f"- {item}" for item in criteria)
+    return f"Your response must satisfy all of the following criteria:\n{bullets}"
+
+
+class JudgeVerdict(BaseModel):
+    """Structured verdict returned by the judge chat client."""
+
+    answered: bool = Field(
+        description=(
+            "True if the agent has fully addressed the original request and it adheres to the other "
+            "judging standards, otherwise False."
+        ),
+    )
+    reasoning: str = Field(
+        default="",
+        description="Brief justification for the verdict.",
+    )
+
+
+# Default iteration cap applied when ``max_iterations`` is not provided. Loops are bounded by
+# default to guard against runaway re-invocation; pass ``max_iterations=None`` explicitly to opt
+# into an unbounded loop.
+DEFAULT_MAX_ITERATIONS = 10
+
+# Default iteration cap for judge-driven loops. LLM-judged loops are costly and probabilistic, so
+# they are bounded by a smaller default. Pass ``max_iterations=None`` explicitly to opt into an
+# unbounded judge loop.
+DEFAULT_JUDGE_MAX_ITERATIONS = 5
+
+
+# A callable invoked between iterations. It always receives the loop keyword arguments
+# (``iteration``, ``last_result``, ``messages``, ``original_messages``, ``session``, ``agent``,
+# ``progress``, ``feedback``). Callers declare only the keywords they need plus ``**kwargs`` to
+# ignore the rest. ``should_continue`` may return a plain ``bool`` (continue/stop) or a
+# ``(bool, str | None)`` tuple whose second item is feedback surfaced to the ``next_message`` and
+# ``record_feedback`` callables via the ``feedback`` keyword argument.
+ShouldContinueResult: TypeAlias = "bool | tuple[bool, str | None]"
+ShouldContinueCallable = Callable[..., "ShouldContinueResult | Awaitable[ShouldContinueResult]"]
+NextMessageCallable = Callable[..., "AgentRunInputs | Awaitable[AgentRunInputs | None] | None"]
+
+# A callable invoked once per work iteration to capture a progress-log entry from that iteration. It
+# receives the loop keyword arguments and returns a string entry (appended to the log) or ``None``
+# (record nothing for that iteration).
+FeedbackCallable = Callable[..., "str | Awaitable[str | None] | None"]
+
+
+async def _maybe_await(value: Any) -> Any:
+    """Await ``value`` if it is awaitable, otherwise return it as-is."""
+    if inspect.isawaitable(value):
+        return await value
+    return value
+
+
+def _build_judge_condition(
+    judge_client: SupportsChatGetResponse,
+    instructions: str,
+) -> tuple[ShouldContinueCallable, NextMessageCallable]:
+    """Build the ``should_continue`` predicate and ``next_message`` callable for a judge loop.
+
+    The judge is called directly (no agent tools, session, or middleware) with fresh messages, so
+    the loop's evaluation cannot recurse back through the agent pipeline. The original input messages
+    are forwarded verbatim (rather than collapsed to text) so multi-modal requests are preserved. The
+    judge is asked for a :class:`JudgeVerdict` structured output; if the client does not honor
+    structured output the verdict falls back to the explicit, non-overlapping ``VERDICT: DONE`` /
+    ``VERDICT: MORE`` markers (``MORE`` wins, keeping the loop running, when the marker is ambiguous
+    or absent).
+
+    The predicate returns a ``(continue, reasoning)`` tuple; the loop surfaces that ``reasoning`` to
+    the next-message callable as the ``feedback`` keyword argument, which feeds it back to the agent
+    so it knows *why* its previous answer was judged incomplete.
+    """
+
+    async def _judge(
+        *, last_result: AgentResponse, original_messages: list[Message], **kwargs: Any
+    ) -> tuple[bool, str | None]:
+        judge_messages = [
+            Message(role="system", contents=[instructions]),
+            Message(
+                role="user",
+                contents=["Evaluate the agent's work. The user's original request follows:"],
+            ),
+            *original_messages,
+            Message(role="user", contents=["The agent's latest response was:"]),
+            *last_result.messages,
+            Message(role="user", contents=["Has the original request been fully addressed?"]),
+        ]
+        response = await judge_client.get_response(judge_messages, options={"response_format": JudgeVerdict})
+        verdict = response.value
+        if isinstance(verdict, JudgeVerdict):
+            answered = verdict.answered
+            reasoning = verdict.reasoning
+        else:
+            # Fallback for clients that do not honor structured output: look for the explicit,
+            # non-overlapping verdict markers. ``FAIL`` (more work needed) takes precedence so an
+            # ambiguous or marker-less reply keeps looping rather than stopping on an incomplete
+            # answer.
+            text = response.text.upper()
+            # ``MORE`` (more work needed) takes precedence so an ambiguous reply keeps looping.
+            answered = False if JUDGE_VERDICT_MORE in text else JUDGE_VERDICT_DONE in text
+            reasoning = response.text.strip()
+        # Continue looping while the request is not yet answered, surfacing the reasoning as feedback.
+        return (not answered), (reasoning or None)
+
+    def _next_message(*, feedback: str | None = None, **kwargs: Any) -> AgentRunInputs:
+        # Feed the judge's reasoning back to the agent so the next iteration addresses the gap.
+        if feedback:
+            return (
+                "An evaluator reviewed your previous response and judged that it does not yet fully "
+                f"address the original request.\n\nEvaluator feedback: {feedback}\n\n"
+                "Revise and continue so the original request is fully addressed."
+            )
+        return DEFAULT_NEXT_MESSAGE
+
+    return _judge, _next_message
+
+
+@experimental(feature_id=ExperimentalFeature.HARNESS)
+class AgentLoopMiddleware(AgentMiddleware):
+    """Re-run an agent in a loop until a criterion is met (or never).
+
+    This middleware repeatedly invokes the wrapped agent. After each run it decides whether to run
+    again based on ``should_continue`` and ``max_iterations``, and uses ``next_message`` to build
+    the input for the next iteration. Use :meth:`with_judge` to drive the loop with a chat-client
+    judge instead of a hand-written predicate.
+
+    By default a non-streaming run returns an aggregated :class:`~agent_framework.AgentResponse`
+    containing every iteration's messages plus the injected ``next_message`` "nudge" messages (set
+    ``return_final_only=True`` to return only the last iteration's response). Streaming runs always
+    yield each iteration's updates and emit the injected nudge messages as ``user`` updates between
+    iterations.
+
+    The ``should_continue`` and ``next_message`` callables are invoked with keyword arguments, so a
+    caller only needs to declare the ones it uses plus ``**kwargs``. The keywords are:
+
+    - ``iteration`` (int): the number of completed runs so far (1-based after the first run).
+    - ``last_result`` (AgentResponse): the result of the iteration that just completed.
+    - ``messages`` (list[Message]): the messages used for the iteration that just completed.
+    - ``original_messages`` (list[Message]): the input used for the first iteration.
+    - ``session`` (AgentSession | None): the active session, used by the provider helpers.
+    - ``agent``: the agent being looped.
+    - ``progress`` (list[str]): the feedback log accumulated so far (see ``record_feedback``).
+    - ``feedback`` (str | None): the feedback string returned by ``should_continue`` for this
+      iteration (``None`` when it returned a plain bool). ``should_continue`` may return either a
+      ``bool`` or a ``(bool, str | None)`` tuple; the string is surfaced here so ``next_message``
+      and ``record_feedback`` can reference it.
+
+    Examples:
+        .. code-block:: python
+
+            from agent_framework import Agent, AgentResponse
+            from agent_framework._harness._loop import AgentLoopMiddleware
+
+
+            async def should_continue(*, iteration: int, last_result: AgentResponse, **kwargs) -> bool:
+                return iteration < 3 and "DONE" not in last_result.text
+
+
+            agent = Agent(client=client, middleware=[AgentLoopMiddleware(should_continue)])
+
+    Note:
+        ``max_iterations`` acts as a safety cap and defaults to ``DEFAULT_MAX_ITERATIONS`` (10). Pass
+        an explicit ``None`` to make the loop unbounded, in which case it relies entirely on
+        ``should_continue`` to stop, so make sure the predicate can eventually return ``False``.
+    """
+
+    def __init__(
+        self,
+        should_continue: ShouldContinueCallable,
+        *,
+        max_iterations: int | None = DEFAULT_MAX_ITERATIONS,
+        next_message: NextMessageCallable | None = None,
+        record_feedback: FeedbackCallable | None = None,
+        inject_progress: bool = True,
+        fresh_context: bool = False,
+        return_final_only: bool = False,
+        additional_instructions: str | None = None,
+    ) -> None:
+        """Initialize the agent loop middleware.
+
+        Args:
+            should_continue: Predicate that decides whether to run the agent again. May be sync or
+                async and is called with the loop keyword arguments (``iteration``, ``last_result``,
+                ``messages``, ``original_messages``, ``session``, ``agent``, ``progress``, and
+                ``feedback`` -- see the class docstring for what each one carries; declare only the
+                ones you need plus ``**kwargs``). Return ``True``/``False`` to
+                continue/stop, or a ``(bool, str | None)`` tuple to also provide feedback; the
+                feedback string is surfaced to the ``next_message`` and ``record_feedback`` callables
+                via the ``feedback`` keyword argument. To loop on a chat-client judge instead, build
+                the middleware via :meth:`with_judge`.
+
+        Keyword Args:
+            max_iterations: Maximum number of agent runs, used as a safety cap. Defaults to
+                ``DEFAULT_MAX_ITERATIONS`` (10); pass an explicit ``None`` for an unbounded loop, or
+                a positive integer to set a custom cap. (The :meth:`with_judge` factory uses
+                ``DEFAULT_JUDGE_MAX_ITERATIONS`` (5) as its default instead.)
+            next_message: Callable that produces the input for the next iteration, called with the
+                loop keyword arguments. Defaults to a short "continue" nudge. Returning ``None``
+                reuses the previous iteration's messages verbatim (in which case the progress log is
+                *not* injected; see ``inject_progress``).
+            record_feedback: Optional callable invoked once per work iteration to capture a feedback
+                entry. Called as ``record_feedback(**loop_kwargs)`` and returns a
+                string entry appended to the progress log, or ``None`` to record nothing for that
+                iteration. When not provided, the iteration's response text (``last_result.text``) is
+                recorded instead.                 The accumulated log is exposed to every callback via the
+                ``progress`` loop keyword argument. For production loops prefer a ``record_feedback``
+                that returns a terse summary rather than relying on the full response text.
+            inject_progress: When ``True`` (default), the accumulated progress log is injected into
+                the next iteration's input as a single ``user`` message ("Progress so far: ..."). To
+                avoid duplication, only the most recent entry is injected when a session is attached
+                (the session already retains earlier turns); the full log is injected when there is
+                no session or ``fresh_context`` is set. When ``False`` the log is only exposed via the
+                ``progress`` loop keyword argument and never injected automatically.
+            fresh_context: When ``True``, each iteration starts from a clean context: ``context``
+                messages are reset to the original input messages (plus the injected progress log)
+                instead of accumulating the prior conversation. When a session is attached, the
+                session is snapshotted once before the loop and restored to that pre-loop baseline
+                before each subsequent iteration, so the local transcript and any service-side
+                conversation id are reset too and the agent does not re-read the accumulated history.
+                In-loop working-state mutations are discarded; pre-loop state is preserved; continuity
+                is carried only by the progress log.
+            return_final_only: Controls what a non-streaming run returns. When ``False`` (default),
+                the returned :class:`~agent_framework.AgentResponse` aggregates every iteration: each
+                iteration's response messages plus the injected ``next_message`` "nudge" messages
+                (as ``user`` messages), so the caller sees the full back-and-forth. When ``True``,
+                only the final iteration's :class:`~agent_framework.AgentResponse` is returned. This
+                flag has no effect on streaming runs (the stream cannot know in advance which
+                iteration is last); streaming always yields each iteration's updates and injects the
+                ``next_message`` messages as ``user`` updates between iterations.
+            additional_instructions: Optional extra instruction injected as a ``system`` message
+                ahead of the input messages before the agent runs. It becomes part of the original
+                messages, so it is preserved across ``fresh_context`` resets and (with a session)
+                persists server-side across iterations. Used by :meth:`with_judge` to tell the agent
+                about the criteria its response must satisfy, but available to any loop.
+
+        Raises:
+            ValueError: If ``max_iterations`` is not ``None`` and is less than 1.
+        """
+        if max_iterations is not None and max_iterations < 1:
+            raise ValueError("max_iterations must be None or a positive integer (>= 1).")
+
+        self.max_iterations: int | None = max_iterations
+        self.should_continue: ShouldContinueCallable = should_continue
+        self.next_message = next_message
+        self.record_feedback = record_feedback
+        self.inject_progress = inject_progress
+        self.fresh_context = fresh_context
+        self.return_final_only = return_final_only
+        self.additional_instructions = additional_instructions
+
+    @classmethod
+    def with_judge(
+        cls,
+        judge_client: SupportsChatGetResponse,
+        *,
+        criteria: Sequence[str] | None = None,
+        instructions: str | None = None,
+        max_iterations: int | None = DEFAULT_JUDGE_MAX_ITERATIONS,
+        next_message: NextMessageCallable | None = None,
+        fresh_context: bool = False,
+    ) -> Self:
+        """Create a loop that continues until a judge chat client decides the request was answered.
+
+        Convenience factory for the judge pattern: ``judge_client`` is queried with a
+        :class:`JudgeVerdict` structured-output response after each iteration and the loop continues
+        while the request is *not* answered. The judge's ``reasoning`` is fed back to the agent as
+        the next iteration's input (unless a custom ``next_message`` is provided), so the agent knows
+        why its previous answer was judged incomplete. See :meth:`__init__` for the full meaning of
+        each argument.
+
+        Args:
+            judge_client: Chat client used to judge whether the original request was answered.
+
+        Keyword Args:
+            criteria: Optional list of criteria the response must satisfy. When provided, they are
+                (1) injected as an extra ``system`` instruction for the agent before it runs (via
+                ``additional_instructions``) and (2) rendered into the judge instructions wherever
+                the ``{{criteria}}`` placeholder appears (``CRITERIA_PLACEHOLDER``).
+            instructions: Optional system instructions for the judge. Defaults to
+                ``DEFAULT_JUDGE_INSTRUCTIONS``. May contain the ``{{criteria}}`` placeholder, which
+                is replaced with the rendered ``criteria`` (or removed when no criteria are given).
+            max_iterations: Maximum number of agent runs. Defaults to
+                ``DEFAULT_JUDGE_MAX_ITERATIONS`` (5); pass ``None`` for unbounded, or a positive
+                integer to set a custom cap.
+            next_message: Callable that produces the next iteration's input. Defaults to one that
+                relays the judge's ``reasoning`` back to the agent.
+            fresh_context: When ``True``, each iteration restarts from the original input messages
+                (plus the injected progress log and judge feedback) instead of accumulating the prior
+                conversation; an attached session is snapshotted before the loop and restored to that
+                baseline between iterations. See :meth:`__init__` for the full semantics. Defaults to
+                ``False``.
+        """
+        judge_instructions = (instructions or DEFAULT_JUDGE_INSTRUCTIONS).replace(
+            CRITERIA_PLACEHOLDER, _render_criteria_block(criteria)
+        )
+        should_continue, judge_next_message = _build_judge_condition(judge_client, judge_instructions)
+        return cls(
+            should_continue=should_continue,
+            max_iterations=max_iterations,
+            next_message=next_message or judge_next_message,
+            fresh_context=fresh_context,
+            additional_instructions=_criteria_agent_instruction(criteria) if criteria else None,
+        )
+
+    async def process(
+        self,
+        context: AgentContext,
+        call_next: Callable[[], Awaitable[None]],
+    ) -> None:
+        """Run the wrapped agent in a loop."""
+        if self.additional_instructions is not None:
+            # Inject the extra instruction as a system message ahead of the input so it is present
+            # on every iteration and preserved across fresh_context resets (which restart from
+            # ``original_messages``).
+            context.messages = [
+                Message(role="system", contents=[self.additional_instructions]),
+                *context.messages,
+            ]
+        original_messages = list(context.messages)
+        # For a truly fresh context per iteration the session must also be reset, otherwise the
+        # next run reloads the local transcript or re-threads the service-side conversation and the
+        # model still sees the accumulated history. Snapshot the session once here (the pre-loop
+        # baseline) and restore it before each subsequent iteration so every pass starts clean.
+        snapshot = context.session.to_dict() if self.fresh_context and context.session is not None else None
+        if context.stream:
+            self._process_streaming(context, call_next, original_messages, snapshot)
+        else:
+            await self._process_non_streaming(context, call_next, original_messages, snapshot)
+
+    @staticmethod
+    def _restore_session(session: Any, snapshot: dict[str, Any]) -> None:
+        """Restore a session in place to a previously captured ``to_dict()`` snapshot.
+
+        Re-hydrates the snapshot via :meth:`AgentSession.from_dict` and copies the mutable fields
+        (``service_session_id`` and ``state``) back onto the live ``session`` instance, so any
+        reference held by the agent/context observes the reset. ``session_id`` is preserved (the
+        snapshot carries the same id). A fresh ``from_dict`` is built on every call so repeated
+        restores from one snapshot do not alias the same state dict.
+        """
+        from .._sessions import AgentSession
+
+        restored = AgentSession.from_dict(snapshot)
+        session.service_session_id = restored.service_session_id
+        session.state = restored.state
+
+    async def _process_non_streaming(
+        self,
+        context: AgentContext,
+        call_next: Callable[[], Awaitable[None]],
+        original_messages: list[Message],
+        snapshot: dict[str, Any] | None,
+    ) -> None:
+        iteration = 0
+        work_iterations = 0
+        progress: list[str] = []
+        # Aggregated transcript across iterations: each iteration's response messages plus the
+        # injected "nudge" messages, used to build the combined response when return_final_only=False.
+        aggregated: list[Message] = []
+        aggregated_usage: UsageDetails | None = None
+        final_result: AgentResponse | None = None
+        while True:
+            await call_next()
+            iteration += 1
+
+            result = context.result
+            if not isinstance(result, AgentResponse):
+                raise TypeError(
+                    "AgentLoopMiddleware expected an AgentResponse from a non-streaming run, "
+                    f"got {type(result).__name__}."
+                )
+
+            final_result = result
+            aggregated.extend(result.messages)
+            if result.usage_details is not None:
+                aggregated_usage = add_usage_details(aggregated_usage, result.usage_details)
+
+            messages_used = context.messages
+            loop_kwargs = self._build_loop_kwargs(
+                context=context,
+                iteration=iteration,
+                last_result=result,
+                messages_used=messages_used,
+                original_messages=original_messages,
+                progress=progress,
+            )
+
+            work_iterations += 1
+            # Decide whether to stop and capture any feedback from should_continue first, so the
+            # feedback is available to both the progress and next-message callables this iteration.
+            stop, feedback = await self._evaluate_stop(loop_kwargs, work_iterations)
+            loop_kwargs = self._build_loop_kwargs(
+                context=context,
+                iteration=iteration,
+                last_result=result,
+                messages_used=messages_used,
+                original_messages=original_messages,
+                progress=progress,
+                feedback=feedback,
+            )
+            # Capture this iteration's progress entry, then refresh loop_kwargs so the next-message
+            # resolution sees the latest entry.
+            if await self._record_progress(result, loop_kwargs, progress):
+                loop_kwargs = self._build_loop_kwargs(
+                    context=context,
+                    iteration=iteration,
+                    last_result=result,
+                    messages_used=messages_used,
+                    original_messages=original_messages,
+                    progress=progress,
+                    feedback=feedback,
+                )
+            if stop:
+                break
+            if snapshot is not None and context.session is not None:
+                # Reset the session to the pre-loop baseline so the next run starts fresh; only the
+                # progress log (injected by _resolve_next_message) carries continuity forward.
+                self._restore_session(context.session, snapshot)
+            next_messages = await self._resolve_next_message(loop_kwargs, messages_used, original_messages)
+            context.messages = next_messages
+            aggregated.extend(next_messages)
+
+        if not self.return_final_only:
+            context.result = self._aggregate_response(final_result, aggregated, aggregated_usage)
+
+    def _process_streaming(
+        self,
+        context: AgentContext,
+        call_next: Callable[[], Awaitable[None]],
+        original_messages: list[Message],
+        snapshot: dict[str, Any] | None,
+    ) -> None:
+        # Holds the last iteration's final response so the outer stream's finalizer can return it
+        # rather than an aggregate of every iteration.
+        holder: dict[str, AgentResponse | None] = {"final": None}
+
+        async def _generator() -> Any:
+            iteration = 0
+            work_iterations = 0
+            progress: list[str] = []
+            while True:
+                try:
+                    await call_next()
+                    inner = context.result
+                    if not isinstance(inner, ResponseStream):
+                        raise TypeError(
+                            "AgentLoopMiddleware expected a ResponseStream from a streaming run, "
+                            f"got {type(inner).__name__}."
+                        )
+
+                    async for update in inner:
+                        yield update
+
+                    holder["final"] = await inner.get_final_response()
+                except MiddlewareTermination:
+                    # The pipeline's MiddlewareTermination suppression is no longer active once
+                    # process() has returned (the stream is consumed lazily), so a termination
+                    # raised by a downstream middleware or during stream consumption surfaces here.
+                    # Stop cleanly and keep whatever final response we have from a prior iteration.
+                    return
+
+                iteration += 1
+
+                messages_used = context.messages
+                final = holder["final"]
+                loop_kwargs = self._build_loop_kwargs(
+                    context=context,
+                    iteration=iteration,
+                    last_result=final,
+                    messages_used=messages_used,
+                    original_messages=original_messages,
+                    progress=progress,
+                )
+
+                work_iterations += 1
+                # Decide whether to stop and capture any feedback from should_continue first, so the
+                # feedback is available to both the progress and next-message callables this iteration.
+                stop, feedback = await self._evaluate_stop(loop_kwargs, work_iterations)
+                loop_kwargs = self._build_loop_kwargs(
+                    context=context,
+                    iteration=iteration,
+                    last_result=final,
+                    messages_used=messages_used,
+                    original_messages=original_messages,
+                    progress=progress,
+                    feedback=feedback,
+                )
+                if await self._record_progress(final, loop_kwargs, progress):
+                    loop_kwargs = self._build_loop_kwargs(
+                        context=context,
+                        iteration=iteration,
+                        last_result=final,
+                        messages_used=messages_used,
+                        original_messages=original_messages,
+                        progress=progress,
+                        feedback=feedback,
+                    )
+                if stop:
+                    return
+                if snapshot is not None and context.session is not None:
+                    # Reset the session to the pre-loop baseline before the next run. The final
+                    # response was already awaited above, so the service-side conversation id has
+                    # been propagated and is safe to discard here.
+                    self._restore_session(context.session, snapshot)
+                next_messages = await self._resolve_next_message(loop_kwargs, messages_used, original_messages)
+                context.messages = next_messages
+                # Surface the injected "nudge" messages in the stream so consumers see the user
+                # turns that drive each subsequent iteration (the equivalent of the aggregated
+                # transcript that non-streaming runs return).
+                for message in next_messages:
+                    yield self._message_to_update(message)
+
+        def _finalize(updates: Sequence[AgentResponseUpdate]) -> AgentResponse:
+            if holder["final"] is not None:
+                return holder["final"]
+            return AgentResponse.from_updates(updates)
+
+        context.result = ResponseStream(_generator(), finalizer=_finalize)
+
+    def _build_loop_kwargs(
+        self,
+        *,
+        context: AgentContext,
+        iteration: int,
+        last_result: AgentResponse | None,
+        messages_used: list[Message],
+        original_messages: list[Message],
+        progress: list[str],
+        feedback: str | None = None,
+    ) -> dict[str, Any]:
+        return {
+            "iteration": iteration,
+            "last_result": last_result,
+            "messages": messages_used,
+            "original_messages": original_messages,
+            "session": context.session,
+            "agent": context.agent,
+            # A copy so user callbacks cannot mutate the loop's internal progress log.
+            "progress": list(progress),
+            # Feedback returned by ``should_continue`` for this iteration (``None`` if it returned a
+            # plain bool, or the stop was decided by ``max_iterations``).
+            "feedback": feedback,
+        }
+
+    async def _record_progress(
+        self,
+        last_result: AgentResponse | None,
+        loop_kwargs: dict[str, Any],
+        progress: list[str],
+    ) -> bool:
+        """Capture this iteration's feedback into ``progress``. Returns ``True`` if an entry was added."""
+        if self.record_feedback is not None:
+            entry = await _maybe_await(self.record_feedback(**loop_kwargs))
+        else:
+            entry = last_result.text.strip() if last_result is not None else None
+        if entry:
+            progress.append(entry)
+            return True
+        return False
+
+    async def _evaluate_stop(self, loop_kwargs: dict[str, Any], work_iterations: int) -> tuple[bool, str | None]:
+        """Decide whether the loop should stop, returning ``(stop, feedback)``.
+
+        ``max_iterations`` is a safety cap that short-circuits before ``should_continue`` is
+        evaluated (so an expensive predicate/judge is not called once the cap has fired). Any
+        feedback returned by ``should_continue`` is propagated so the progress and next-message
+        callables can reference it.
+        """
+        if self.max_iterations is not None and work_iterations >= self.max_iterations:
+            return True, None
+        keep_going, feedback = await self._should_continue(loop_kwargs)
+        return (not keep_going), feedback
+
+    async def _should_continue(self, loop_kwargs: dict[str, Any]) -> tuple[bool, str | None]:
+        """Evaluate the predicate, normalizing its result to ``(continue, feedback)``."""
+        result = await _maybe_await(self.should_continue(**loop_kwargs))
+        return (bool(result[0]), result[1]) if isinstance(result, tuple) else (bool(result), None)  # type: ignore
+
+    @staticmethod
+    def _message_to_update(message: Message) -> AgentResponseUpdate:
+        """Wrap an injected loop message as a streaming update so consumers see it inline."""
+        return AgentResponseUpdate(
+            contents=message.contents,
+            role=message.role,
+            author_name=message.author_name,
+            message_id=message.message_id,
+        )
+
+    @staticmethod
+    def _aggregate_response(
+        final: AgentResponse,
+        messages: list[Message],
+        usage: UsageDetails | None,
+    ) -> AgentResponse:
+        """Build a combined response carrying every iteration's messages and summed usage.
+
+        Metadata (``response_id``, structured ``value``, etc.) is taken from the final iteration; the
+        structured value is passed through pre-parsed so it is not re-derived from the aggregated text.
+        """
+        return AgentResponse(
+            messages=messages,
+            response_id=final.response_id,
+            agent_id=final.agent_id,
+            created_at=final.created_at,
+            finish_reason=final.finish_reason,  # pyright: ignore[reportArgumentType]
+            usage_details=usage,
+            value=final.value,
+            additional_properties=dict(final.additional_properties) if final.additional_properties else None,
+            raw_representation=final.raw_representation,
+        )
+
+    @staticmethod
+    def _render_progress(entries: list[str]) -> Message:
+        """Format progress-log entries into a single ``user`` message."""
+        body = "\n".join(f"- {entry}" for entry in entries)
+        return Message(role="user", contents=[f"Progress so far:\n{body}"])
+
+    async def _resolve_next_message(
+        self,
+        loop_kwargs: dict[str, Any],
+        messages_used: list[Message],
+        original_messages: list[Message],
+    ) -> list[Message]:
+        # Compute the base next input. A ``next_message`` callable returning None requests a verbatim
+        # reuse of the previous messages (no progress injection); in fresh-context mode that escape
+        # hatch does not apply, so fall back to the default nudge instead.
+        if self.next_message is None:
+            next_msgs = normalize_messages(DEFAULT_NEXT_MESSAGE)
+        else:
+            next_input = await _maybe_await(self.next_message(**loop_kwargs))
+            if next_input is None:
+                if not self.fresh_context:
+                    return list(messages_used)
+                next_msgs = normalize_messages(DEFAULT_NEXT_MESSAGE)
+            else:
+                next_msgs = normalize_messages(next_input)
+
+        progress: list[str] = loop_kwargs.get("progress") or []
+        session = loop_kwargs.get("session")
+        progress_msg: Message | None = None
+        if self.inject_progress and progress:
+            # With a session the earlier entries are already retained in the conversation, so only
+            # the latest entry is injected to avoid duplication. Otherwise inject the full log.
+            entries = progress if (session is None or self.fresh_context) else progress[-1:]
+            progress_msg = self._render_progress(entries)
+
+        if self.fresh_context:
+            result = list(original_messages)
+            if progress_msg is not None:
+                result.append(progress_msg)
+            result.extend(next_msgs)
+            return result
+
+        if progress_msg is not None:
+            return [progress_msg, *next_msgs]
+        return list(next_msgs)
+
+
+def todos_remaining(provider: Any) -> ShouldContinueCallable:
+    """Build a ``should_continue`` predicate that loops while a ``TodoProvider`` has open items.
+
+    Args:
+        provider: A :class:`~agent_framework.TodoProvider` attached to the same session as the loop.
+
+    Returns:
+        A predicate suitable for :class:`AgentLoopMiddleware`'s ``should_continue`` argument.
+    """
+
+    async def _should_continue(*, session: Any = None, **kwargs: Any) -> bool:
+        if session is None:
+            return False
+        items = await provider.store.load_items(session, source_id=provider.source_id)
+        return any(not item.is_complete for item in items)
+
+    return _should_continue
+
+
+def background_tasks_running(provider: Any) -> ShouldContinueCallable:
+    """Build a ``should_continue`` predicate that loops while a ``BackgroundAgentsProvider`` is busy.
+
+    The predicate inspects the provider's persisted task state and continues while any task is still
+    marked as running. Pair it with ``max_iterations`` so the loop is guaranteed to stop even if a
+    task's persisted status is never refreshed.
+
+    Args:
+        provider: A :class:`~agent_framework.BackgroundAgentsProvider` attached to the same session
+            as the loop.
+
+    Returns:
+        A predicate suitable for :class:`AgentLoopMiddleware`'s ``should_continue`` argument.
+    """
+    from ._background_agents import BackgroundTaskInfo, BackgroundTaskStatus
+
+    def _should_continue(*, session: Any = None, **kwargs: Any) -> bool:
+        if session is None:
+            return False
+        state = session.state.get(provider.source_id)
+        if not state:
+            return False
+        return any(
+            BackgroundTaskInfo.from_dict(task).status == BackgroundTaskStatus.RUNNING for task in state.get("tasks", [])
+        )
+
+    return _should_continue

python/packages/core/agent_framework/_types.py

@@ -400,12 +400,18 @@ class UsageDetails(TypedDict, total=False, extra_items=int):  # type: ignore[cal
         input_token_count: The number of input tokens used.
         output_token_count: The number of output tokens generated.
         total_token_count: The total number of tokens (input + output).
+        cache_creation_input_token_count: The number of input tokens written to a provider-managed cache.
+        cache_read_input_token_count: The number of input tokens served from a provider-managed cache.
+        reasoning_output_token_count: The number of output tokens used for reasoning.
 
     """
 
     input_token_count: int | None
     output_token_count: int | None
     total_token_count: int | None
+    cache_creation_input_token_count: int | None
+    cache_read_input_token_count: int | None
+    reasoning_output_token_count: int | None
 
 
 def add_usage_details(usage1: UsageDetails | None, usage2: UsageDetails | None) -> UsageDetails:

python/packages/core/agent_framework/ag_ui/__init__.py

@@ -11,6 +11,10 @@ Supported classes and functions:
 - AGUIChatClient
 - AGUIEventConverter
 - AGUIHttpService
+- AGUIThreadSnapshot
+- AGUIThreadSnapshotStore
+- InMemoryAGUIThreadSnapshotStore
+- SnapshotScopeResolver
 - add_agent_framework_fastapi_endpoint
 - state_update
 - __version__
@@ -28,6 +32,10 @@ _IMPORTS = [
     "AGUIChatClient",
     "AGUIEventConverter",
     "AGUIHttpService",
+    "AGUIThreadSnapshot",
+    "AGUIThreadSnapshotStore",
+    "InMemoryAGUIThreadSnapshotStore",
+    "SnapshotScopeResolver",
     "state_update",
     "__version__",
 ]

python/packages/core/agent_framework/ag_ui/__init__.pyi

@@ -6,6 +6,10 @@ from agent_framework_ag_ui import (
     AGUIChatClient,
     AGUIEventConverter,
     AGUIHttpService,
+    AGUIThreadSnapshot,
+    AGUIThreadSnapshotStore,
+    InMemoryAGUIThreadSnapshotStore,
+    SnapshotScopeResolver,
     __version__,
     add_agent_framework_fastapi_endpoint,
     state_update,
@@ -15,8 +19,12 @@ __all__ = [
     "AGUIChatClient",
     "AGUIEventConverter",
     "AGUIHttpService",
+    "AGUIThreadSnapshot",
+    "AGUIThreadSnapshotStore",
     "AgentFrameworkAgent",
     "AgentFrameworkWorkflow",
+    "InMemoryAGUIThreadSnapshotStore",
+    "SnapshotScopeResolver",
     "__version__",
     "add_agent_framework_fastapi_endpoint",
     "state_update",

python/packages/core/agent_framework/observability.py

@@ -201,6 +201,9 @@ class OtelAttr(str, Enum):
     # Usage attributes
     INPUT_TOKENS = "gen_ai.usage.input_tokens"
     OUTPUT_TOKENS = "gen_ai.usage.output_tokens"
+    CACHE_CREATION_INPUT_TOKENS = "gen_ai.usage.cache_creation.input_tokens"
+    CACHE_READ_INPUT_TOKENS = "gen_ai.usage.cache_read.input_tokens"
+    REASONING_OUTPUT_TOKENS = "gen_ai.usage.reasoning.output_tokens"
     # Tool attributes
     TOOL_CALL_ID = "gen_ai.tool.call.id"
     TOOL_DESCRIPTION = "gen_ai.tool.description"
@@ -327,6 +330,20 @@ FINISH_REASON_MAP = {
     "tool_calls": "tool_call",
     "length": "length",
 }
+USAGE_DETAIL_TO_OTEL_ATTR: Final[tuple[tuple[str, OtelAttr], ...]] = (
+    ("input_token_count", OtelAttr.INPUT_TOKENS),
+    ("output_token_count", OtelAttr.OUTPUT_TOKENS),
+    ("cache_creation_input_token_count", OtelAttr.CACHE_CREATION_INPUT_TOKENS),
+    ("cache_read_input_token_count", OtelAttr.CACHE_READ_INPUT_TOKENS),
+    ("reasoning_output_token_count", OtelAttr.REASONING_OUTPUT_TOKENS),
+    ("anthropic.cache_creation_input_tokens", OtelAttr.CACHE_CREATION_INPUT_TOKENS),
+    ("anthropic.cache_read_input_tokens", OtelAttr.CACHE_READ_INPUT_TOKENS),
+    ("openai.cached_input_tokens", OtelAttr.CACHE_READ_INPUT_TOKENS),
+    ("prompt/cached_tokens", OtelAttr.CACHE_READ_INPUT_TOKENS),
+    ("openai.reasoning_tokens", OtelAttr.REASONING_OUTPUT_TOKENS),
+    ("completion/reasoning_tokens", OtelAttr.REASONING_OUTPUT_TOKENS),
+    ("reasoning_tokens", OtelAttr.REASONING_OUTPUT_TOKENS),
+)
 
 
 # region Telemetry utils
@@ -2350,12 +2367,16 @@ def _apply_accumulated_usage(attributes: dict[str, Any], captured_fields: set[st
     accumulated = INNER_ACCUMULATED_USAGE.get()
     if not accumulated:
         return
-    input_tokens = accumulated.get("input_token_count")
-    if input_tokens:
-        attributes[OtelAttr.INPUT_TOKENS] = input_tokens
-    output_tokens = accumulated.get("output_token_count")
-    if output_tokens:
-        attributes[OtelAttr.OUTPUT_TOKENS] = output_tokens
+    _apply_usage_attributes(attributes, accumulated)
+
+
+def _apply_usage_attributes(attributes: dict[str, Any], usage: Mapping[str, Any]) -> None:
+    """Apply known usage details as standard OTel GenAI attributes."""
+    for usage_key, otel_attr in USAGE_DETAIL_TO_OTEL_ATTR:
+        value = usage.get(usage_key)
+        if value is None or isinstance(value, bool) or not isinstance(value, int):
+            continue
+        attributes.setdefault(otel_attr, value)
 
 
 def _get_response_attributes(
@@ -2378,12 +2399,7 @@ def _get_response_attributes(
     if model := getattr(response, "model", None):
         attributes[OtelAttr.RESPONSE_MODEL] = model
     if capture_usage and (usage := response.usage_details):
-        input_tokens = usage.get("input_token_count")
-        if input_tokens:
-            attributes[OtelAttr.INPUT_TOKENS] = input_tokens
-        output_tokens = usage.get("output_token_count")
-        if output_tokens:
-            attributes[OtelAttr.OUTPUT_TOKENS] = output_tokens
+        _apply_usage_attributes(attributes, usage)
     return attributes
 
 
@@ -2407,9 +2423,9 @@ def _capture_response(
     """Set the response for a given span."""
     span.set_attributes(attributes)
     attrs: dict[str, Any] = {k: v for k, v in attributes.items() if k in GEN_AI_METRIC_ATTRIBUTES}
-    if token_usage_histogram and (input_tokens := attributes.get(OtelAttr.INPUT_TOKENS)):
+    if token_usage_histogram and (input_tokens := attributes.get(OtelAttr.INPUT_TOKENS)) is not None:
         token_usage_histogram.record(input_tokens, attributes={**attrs, OtelAttr.T_TYPE: OtelAttr.T_TYPE_INPUT})
-    if token_usage_histogram and (output_tokens := attributes.get(OtelAttr.OUTPUT_TOKENS)):
+    if token_usage_histogram and (output_tokens := attributes.get(OtelAttr.OUTPUT_TOKENS)) is not None:
         token_usage_histogram.record(output_tokens, {**attrs, OtelAttr.T_TYPE: OtelAttr.T_TYPE_OUTPUT})
     if operation_duration_histogram and duration is not None:
         if OtelAttr.ERROR_TYPE in attributes:

python/packages/core/tests/core/test_harness_file_access.py

@@ -158,6 +158,69 @@ async def test_in_memory_store_search_returns_matches_with_snippets() -> None:
     assert {result.file_name for result in results_all} == {"a.md", "notes.txt"}
 
 
+async def test_in_memory_store_search_is_recursive_with_root_relative_names() -> None:
+    """Recursive search should find files at any depth and return root-relative names."""
+    store = InMemoryAgentFileStore()
+    await store.write_file("top.md", "ERROR at top")
+    await store.write_file("reports/q1.md", "ERROR in q1")
+    await store.write_file("reports/2024/q2.md", "ERROR in q2")
+    await store.write_file("reports/2024/data.txt", "ERROR wrong extension")
+
+    # Non-recursive (default) only sees the direct child.
+    direct = await store.search_files("", "error")
+    assert {result.file_name for result in direct} == {"top.md"}
+
+    # Recursive sees every descendant, with store-root-relative file names.
+    recursive = await store.search_files("", "error", recursive=True)
+    assert {result.file_name for result in recursive} == {
+        "top.md",
+        "reports/q1.md",
+        "reports/2024/q2.md",
+        "reports/2024/data.txt",
+    }
+
+    # Subtree scoping via the glob (``*`` crosses ``/`` with fnmatch).
+    scoped = await store.search_files("", "error", "reports/*", recursive=True)
+    assert {result.file_name for result in scoped} == {
+        "reports/q1.md",
+        "reports/2024/q2.md",
+        "reports/2024/data.txt",
+    }
+
+    # Extension glob matches markdown at any depth but not other extensions.
+    markdown = await store.search_files("", "error", "*.md", recursive=True)
+    assert {result.file_name for result in markdown} == {
+        "top.md",
+        "reports/q1.md",
+        "reports/2024/q2.md",
+    }
+
+
+async def test_in_memory_store_list_directories() -> None:
+    """``list_directories`` should return direct child subdirectories only, preserving casing."""
+    store = InMemoryAgentFileStore()
+    await store.write_file("top.md", "x")
+    await store.write_file("Reports/q1.md", "x")
+    await store.write_file("Reports/2024/q2.md", "x")
+    await store.write_file("data/raw.csv", "x")
+
+    assert sorted(await store.list_directories()) == ["Reports", "data"]
+    assert sorted(await store.list_directories("Reports")) == ["2024"]
+    # A directory with no subdirectories returns an empty list.
+    assert await store.list_directories("data") == []
+    # A missing directory returns an empty list.
+    assert await store.list_directories("missing") == []
+
+
+async def test_in_memory_store_list_directories_rejects_traversal() -> None:
+    """``list_directories`` must reject traversal inputs the way ``list_files`` does."""
+    store = InMemoryAgentFileStore()
+    await store.write_file("reports/q1.md", "x")
+    for bad in ("../escape", "/abs/path", ".."):
+        with pytest.raises(ValueError):
+            await store.list_directories(bad)
+
+
 async def test_in_memory_store_search_rejects_invalid_and_oversize_regex() -> None:
     """``search_files`` should surface clean errors for bad regex input."""
     store = InMemoryAgentFileStore()
@@ -267,6 +330,78 @@ async def test_filesystem_store_search_matches_lines_and_filters_globs(tmp_path:
     assert {result.file_name for result in results_all} == {"a.md", "b.txt"}
 
 
+async def test_filesystem_store_search_is_recursive_with_root_relative_names(tmp_path: Path) -> None:
+    """Recursive filesystem search should walk the subtree and return root-relative names."""
+    store = FileSystemAgentFileStore(tmp_path)
+    await store.write_file("top.md", "ERROR at top")
+    await store.write_file("reports/q1.md", "ERROR in q1")
+    await store.write_file("reports/2024/q2.md", "ERROR in q2")
+
+    direct = await store.search_files("", "error")
+    assert {result.file_name for result in direct} == {"top.md"}
+
+    recursive = await store.search_files("", "error", recursive=True)
+    assert {result.file_name for result in recursive} == {
+        "top.md",
+        "reports/q1.md",
+        "reports/2024/q2.md",
+    }
+
+    scoped = await store.search_files("", "error", "reports/*", recursive=True)
+    assert {result.file_name for result in scoped} == {
+        "reports/q1.md",
+        "reports/2024/q2.md",
+    }
+
+
+async def test_filesystem_store_list_directories(tmp_path: Path) -> None:
+    """``list_directories`` should list direct child subdirectories only."""
+    store = FileSystemAgentFileStore(tmp_path)
+    await store.write_file("top.md", "x")
+    await store.write_file("reports/q1.md", "x")
+    await store.write_file("reports/2024/q2.md", "x")
+    await store.write_file("data/raw.csv", "x")
+
+    assert sorted(await store.list_directories()) == ["data", "reports"]
+    assert sorted(await store.list_directories("reports")) == ["2024"]
+    assert await store.list_directories("data") == []
+    assert await store.list_directories("missing") == []
+
+
+async def test_filesystem_store_list_directories_rejects_traversal(tmp_path: Path) -> None:
+    """``list_directories`` is security-critical and must reject paths that escape the root."""
+    store = FileSystemAgentFileStore(tmp_path)
+    await store.write_file("reports/q1.md", "x")
+    for bad in ("../escape", "/etc", "C:/Windows", ".."):
+        with pytest.raises(ValueError):
+            await store.list_directories(bad)
+
+
+async def test_filesystem_store_search_and_list_skip_symlinked_directories(tmp_path: Path) -> None:
+    """Recursive search must not descend into symlinked dirs and ``list_directories`` must exclude them."""
+    target = tmp_path / "outside"
+    target.mkdir()
+    (target / "secret.md").write_text("ERROR outside the root", encoding="utf-8")
+
+    root = tmp_path / "root"
+    root.mkdir()
+    (root / "inside.md").write_text("ERROR inside", encoding="utf-8")
+    link = root / "linked"
+    try:
+        link.symlink_to(target, target_is_directory=True)
+    except (OSError, NotImplementedError) as exc:
+        pytest.skip(f"Symbolic links are not supported in this environment: {exc!r}")
+
+    store = FileSystemAgentFileStore(root)
+
+    # ``list_directories`` excludes the symlinked directory.
+    assert await store.list_directories() == []
+
+    # Recursive search does not follow the symlink out of the root.
+    results = await store.search_files("", "error", recursive=True)
+    assert {result.file_name for result in results} == {"inside.md"}
+
+
 async def test_filesystem_store_search_skips_non_utf8_files(tmp_path: Path) -> None:
     """The filesystem store should silently skip non-UTF-8 files instead of aborting the search."""
     store = FileSystemAgentFileStore(tmp_path)
@@ -303,7 +438,7 @@ def test_filesystem_store_requires_non_empty_root() -> None:
 async def test_file_access_provider_registers_tools_and_instructions(
     chat_client_base: SupportsChatGetResponse,
 ) -> None:
-    """``FileAccessProvider.before_run`` should add the canonical instructions and five tools."""
+    """``FileAccessProvider.before_run`` should add the canonical instructions and six tools."""
     session = AgentSession(session_id="session-1")
     store = InMemoryAgentFileStore()
     provider = FileAccessProvider(store=store)
@@ -321,6 +456,7 @@ async def test_file_access_provider_registers_tools_and_instructions(
         "file_access_read_file",
         "file_access_delete_file",
         "file_access_list_files",
+        "file_access_list_subdirectories",
         "file_access_search_files",
     }
     assert {getattr(tool, "name", None) for tool in tools} >= expected_names
@@ -354,6 +490,7 @@ async def test_file_access_provider_delete_approval_defaults_to_always_require(
         "file_access_save_file",
         "file_access_read_file",
         "file_access_list_files",
+        "file_access_list_subdirectories",
         "file_access_search_files",
     ):
         assert _tool_by_name(tools, name).approval_mode == "never_require"
@@ -396,6 +533,7 @@ async def test_file_access_provider_tools_round_trip_files(
     read_file = _tool_by_name(tools, "file_access_read_file")
     delete_file = _tool_by_name(tools, "file_access_delete_file")
     list_files = _tool_by_name(tools, "file_access_list_files")
+    list_subdirectories = _tool_by_name(tools, "file_access_list_subdirectories")
     search_files = _tool_by_name(tools, "file_access_search_files")
 
     saved = await save_file.invoke(arguments={"file_name": "plan.md", "content": "step 1\nERROR step 2"})
@@ -426,6 +564,15 @@ async def test_file_access_provider_tools_round_trip_files(
     listed_blank = await list_files.invoke(arguments={"directory": "   "})
     assert sorted(json.loads(listed_blank[0].text)) == ["plan.md"]
 
+    # The subdirectory-discovery tool surfaces child directories (not files).
+    listed_dirs = await list_subdirectories.invoke()
+    assert json.loads(listed_dirs[0].text) == ["reports"]
+    listed_dirs_blank = await list_subdirectories.invoke(arguments={"directory": "   "})
+    assert json.loads(listed_dirs_blank[0].text) == ["reports"]
+    # A leaf directory with no child directories returns an empty list.
+    listed_dirs_nested = await list_subdirectories.invoke(arguments={"directory": "reports"})
+    assert json.loads(listed_dirs_nested[0].text) == []
+
     missing = await read_file.invoke(arguments={"file_name": "missing.md"})
     assert "not found" in missing[0].text
 
@@ -434,14 +581,12 @@ async def test_file_access_provider_tools_round_trip_files(
     assert parsed[0]["file_name"] == "plan.md"
     assert parsed[0]["matching_lines"][0]["line"] == "ERROR replaced"
 
-    # The search tool should likewise accept an optional directory argument so
-    # agents can scope a search to a subfolder.
+    # The search tool is recursive from the store root; scope to a subtree using
+    # the glob (``*`` crosses ``/`` with fnmatch). Results use root-relative names.
     await save_file.invoke(arguments={"file_name": "reports/issues.md", "content": "ERROR nested"})
-    scoped = await search_files.invoke(
-        arguments={"regex_pattern": "error", "file_pattern": "*.md", "directory": "reports"}
-    )
+    scoped = await search_files.invoke(arguments={"regex_pattern": "error", "file_pattern": "reports/*"})
     scoped_parsed = json.loads(scoped[0].text)
-    assert [entry["file_name"] for entry in scoped_parsed] == ["issues.md"]
+    assert [entry["file_name"] for entry in scoped_parsed] == ["reports/issues.md"]
 
     deleted = await delete_file.invoke(arguments={"file_name": "plan.md"})
     assert "deleted" in deleted[0].text

python/packages/core/tests/core/test_observability.py

@@ -2154,6 +2154,58 @@ def test_get_response_attributes_with_usage():
     assert result[OtelAttr.OUTPUT_TOKENS] == 50
 
 
+def test_get_response_attributes_with_additional_usage():
+    """Test _get_response_attributes maps additional usage details to OTel attributes."""
+    from unittest.mock import Mock
+
+    from agent_framework.observability import OtelAttr, _get_response_attributes
+
+    response = Mock()
+    response.response_id = None
+    response.finish_reason = None
+    response.raw_representation = None
+    response.usage_details = {
+        "input_token_count": 0,
+        "output_token_count": 50,
+        "cache_creation_input_token_count": 10,
+        "cache_read_input_token_count": 0,
+        "reasoning_output_token_count": 30,
+    }
+
+    attrs = {}
+    result = _get_response_attributes(attrs, response)
+
+    assert result[OtelAttr.INPUT_TOKENS] == 0
+    assert result[OtelAttr.OUTPUT_TOKENS] == 50
+    assert result[OtelAttr.CACHE_CREATION_INPUT_TOKENS] == 10
+    assert result[OtelAttr.CACHE_READ_INPUT_TOKENS] == 0
+    assert result[OtelAttr.REASONING_OUTPUT_TOKENS] == 30
+
+
+def test_get_response_attributes_maps_legacy_usage_keys():
+    """Test _get_response_attributes maps legacy provider usage keys to standard OTel attributes."""
+    from unittest.mock import Mock
+
+    from agent_framework.observability import OtelAttr, _get_response_attributes
+
+    response = Mock()
+    response.response_id = None
+    response.finish_reason = None
+    response.raw_representation = None
+    response.usage_details = {
+        "anthropic.cache_creation_input_tokens": 12,
+        "openai.cached_input_tokens": 0,
+        "completion/reasoning_tokens": 34,
+    }
+
+    attrs = {}
+    result = _get_response_attributes(attrs, response)
+
+    assert result[OtelAttr.CACHE_CREATION_INPUT_TOKENS] == 12
+    assert result[OtelAttr.CACHE_READ_INPUT_TOKENS] == 0
+    assert result[OtelAttr.REASONING_OUTPUT_TOKENS] == 34
+
+
 def test_get_response_attributes_capture_usage_false():
     """Test _get_response_attributes skips usage when capture_usage is False."""
     from unittest.mock import Mock
@@ -2164,13 +2216,22 @@ def test_get_response_attributes_capture_usage_false():
     response.response_id = None
     response.finish_reason = None
     response.raw_representation = None
-    response.usage_details = {"input_token_count": 100, "output_token_count": 50}
+    response.usage_details = {
+        "input_token_count": 100,
+        "output_token_count": 50,
+        "cache_creation_input_token_count": 10,
+        "cache_read_input_token_count": 20,
+        "reasoning_output_token_count": 30,
+    }
 
     attrs = {}
     result = _get_response_attributes(attrs, response, capture_usage=False)
 
     assert OtelAttr.INPUT_TOKENS not in result
     assert OtelAttr.OUTPUT_TOKENS not in result
+    assert OtelAttr.CACHE_CREATION_INPUT_TOKENS not in result
+    assert OtelAttr.CACHE_READ_INPUT_TOKENS not in result
+    assert OtelAttr.REASONING_OUTPUT_TOKENS not in result
 
 
 def test_get_response_attributes_capture_response_id_false():
@@ -2933,6 +2994,23 @@ def test_capture_response(span_exporter: InMemorySpanExporter):
     assert spans[0].attributes.get(OtelAttr.OUTPUT_TOKENS) == 50
 
 
+def test_capture_response_records_zero_token_usage():
+    """Test _capture_response records zero-valued token usage."""
+    from agent_framework.observability import OtelAttr, _capture_response
+
+    span = Mock()
+    token_histogram = Mock()
+    attrs = {
+        OtelAttr.INPUT_TOKENS: 0,
+        OtelAttr.OUTPUT_TOKENS: 0,
+    }
+
+    _capture_response(span=span, attributes=attrs, token_usage_histogram=token_histogram)
+
+    span.set_attributes.assert_called_once_with(attrs)
+    assert token_histogram.record.call_count == 2
+
+
 async def test_layer_ordering_span_sequence_with_function_calling(span_exporter: InMemorySpanExporter):
     """Test that with correct layer ordering, spans appear in the expected sequence.
 
@@ -3937,11 +4015,21 @@ async def test_agent_invoke_span_aggregates_usage_across_tool_calls(span_exporte
                     Content.from_function_call(call_id="call_1", name="get_weather", arguments='{"city": "Seattle"}')
                 ],
             ),
-            usage_details=UsageDetails(input_token_count=2239, output_token_count=192),
+            usage_details=UsageDetails(
+                input_token_count=2239,
+                output_token_count=192,
+                cache_read_input_token_count=100,
+                reasoning_output_token_count=25,
+            ),
         ),
         ChatResponse(
             messages=Message(role="assistant", contents=["The weather in Seattle is sunny."]),
-            usage_details=UsageDetails(input_token_count=2569, output_token_count=99),
+            usage_details=UsageDetails(
+                input_token_count=2569,
+                output_token_count=99,
+                cache_read_input_token_count=200,
+                reasoning_output_token_count=0,
+            ),
         ),
     ]
 
@@ -3965,12 +4053,18 @@ async def test_agent_invoke_span_aggregates_usage_across_tool_calls(span_exporte
     # Individual chat spans retain their own usage
     assert chat_spans[0].attributes.get(OtelAttr.INPUT_TOKENS) == 2239
     assert chat_spans[0].attributes.get(OtelAttr.OUTPUT_TOKENS) == 192
+    assert chat_spans[0].attributes.get(OtelAttr.CACHE_READ_INPUT_TOKENS) == 100
+    assert chat_spans[0].attributes.get(OtelAttr.REASONING_OUTPUT_TOKENS) == 25
     assert chat_spans[1].attributes.get(OtelAttr.INPUT_TOKENS) == 2569
     assert chat_spans[1].attributes.get(OtelAttr.OUTPUT_TOKENS) == 99
+    assert chat_spans[1].attributes.get(OtelAttr.CACHE_READ_INPUT_TOKENS) == 200
+    assert chat_spans[1].attributes.get(OtelAttr.REASONING_OUTPUT_TOKENS) == 0
 
     # The invoke_agent span must report the aggregate across all LLM round-trips
     assert agent_span.attributes.get(OtelAttr.INPUT_TOKENS) == 2239 + 2569
     assert agent_span.attributes.get(OtelAttr.OUTPUT_TOKENS) == 192 + 99
+    assert agent_span.attributes.get(OtelAttr.CACHE_READ_INPUT_TOKENS) == 100 + 200
+    assert agent_span.attributes.get(OtelAttr.REASONING_OUTPUT_TOKENS) == 25
 
 
 @pytest.mark.parametrize("enable_sensitive_data", [False], indirect=True)

python/packages/declarative/agent_framework_declarative/_workflows/__init__.py

@@ -76,12 +76,10 @@ from ._executors_mcp import (
 from ._executors_tools import (
     FUNCTION_TOOL_REGISTRY_KEY,
     TOOL_ACTION_EXECUTORS,
-    TOOL_APPROVAL_STATE_KEY,
     BaseToolExecutor,
     InvokeFunctionToolExecutor,
     ToolApprovalRequest,
     ToolApprovalResponse,
-    ToolApprovalState,
     ToolInvocationResult,
 )
 from ._factory import WorkflowFactory
@@ -111,7 +109,6 @@ __all__ = [
     "HTTP_ACTION_EXECUTORS",
     "MCP_ACTION_EXECUTORS",
     "TOOL_ACTION_EXECUTORS",
-    "TOOL_APPROVAL_STATE_KEY",
     "TOOL_REGISTRY_KEY",
     "ActionComplete",
     "ActionTrigger",
@@ -164,7 +161,6 @@ __all__ = [
     "SetVariableExecutor",
     "ToolApprovalRequest",
     "ToolApprovalResponse",
-    "ToolApprovalState",
     "ToolInvocationResult",
     "WorkflowFactory",
     "WorkflowState",

python/packages/declarative/agent_framework_declarative/_workflows/_declarative_base.py

@@ -63,6 +63,9 @@ logger = logging.getLogger(__name__)
 
 _ENV_REFERENCE_RE = re.compile(r"\bEnv\.([A-Za-z_][A-Za-z0-9_]*)")
 
+# Allowed identifier shape for object-attribute steps in declarative state paths
+_SAFE_PATH_SEGMENT_RE = re.compile(r"^[A-Za-z][A-Za-z0-9_]*$")
+
 
 @dataclass(frozen=True)
 class DeclarativeEnvConfig:
@@ -266,6 +269,9 @@ class DeclarativeWorkflowState:
     - Conversation: Conversation history
     """
 
+    # Sentinel marking "no prior value" for temporary-key bookkeeping.
+    _MISSING: Any = object()
+
     def __init__(self, state: State, env_config: DeclarativeEnvConfig | None = None):
         """Initialize with a State instance.
 
@@ -331,16 +337,21 @@ class DeclarativeWorkflowState:
     def get(self, path: str, default: Any = None) -> Any:
         """Get a value from the state using a dot-notated path.
 
+        Dict-keyed segments may use arbitrary string keys (e.g. UUIDs in
+        ``System.conversations.<id>.messages``). Segments that would resolve
+        via object-attribute access must be valid declarative identifiers
+        (``[A-Za-z][A-Za-z0-9_]*``); other shapes return ``default``.
+
         Args:
             path: Dot-notated path like 'Local.results' or 'Workflow.Inputs.query'
             default: Default value if path doesn't exist
 
         Returns:
-            The value at the path, or default if not found
+            The value at the path, or default if not found or unreachable.
         """
         state_data = self.get_state_data()
         parts = path.split(".")
-        if not parts:
+        if not parts or any(not p for p in parts):
             return default
 
         namespace = parts[0]
@@ -377,10 +388,19 @@ class DeclarativeWorkflowState:
                 obj = obj.get(part, default)  # type: ignore[union-attr]
                 if obj is default:
                     return default
-            elif hasattr(obj, part):  # type: ignore[arg-type]
-                obj = getattr(obj, part)  # type: ignore[arg-type]
             else:
-                return default
+                # Attribute access is only allowed for safe declarative identifiers.
+                if not _SAFE_PATH_SEGMENT_RE.match(part):
+                    logger.warning(
+                        "DeclarativeWorkflowState.get: rejecting attribute segment %r in path %r",
+                        part,
+                        path,
+                    )
+                    return default
+                if hasattr(obj, part):  # type: ignore[arg-type]
+                    obj = getattr(obj, part)  # type: ignore[arg-type]
+                else:
+                    return default
 
         return obj  # type: ignore[return-value]
 
@@ -392,12 +412,14 @@ class DeclarativeWorkflowState:
             value: The value to set
 
         Raises:
-            ValueError: If attempting to set Workflow.Inputs (which is read-only)
+            ValueError: If ``path`` is empty or contains empty segments
+                (e.g. ``"Local."``, ``"Local..foo"``), or if attempting to set
+                ``Workflow.Inputs`` (which is read-only).
         """
         state_data = self.get_state_data()
         parts = path.split(".")
-        if not parts:
-            return
+        if not parts or any(not p for p in parts):
+            raise ValueError(f"Invalid path {path!r}: empty segments are not allowed")
 
         namespace = parts[0]
         remaining = parts[1:]
@@ -453,7 +475,16 @@ class DeclarativeWorkflowState:
         Args:
             path: Dot-notated path to a list
             value: The value to append
+
+        Raises:
+            ValueError: If ``path`` is empty or contains empty segments
+                (e.g. ``"Local."``, ``"Local..foo"``), or if the existing
+                value at ``path`` is not a list.
         """
+        parts = path.split(".")
+        if not parts or any(not p for p in parts):
+            raise ValueError(f"Invalid path {path!r}: empty segments are not allowed")
+
         existing = self.get(path)
         if existing is None:
             self.set(path, [value])
@@ -464,6 +495,15 @@ class DeclarativeWorkflowState:
         else:
             raise ValueError(f"Cannot append to non-list at path '{path}'")
 
+    def _clear_local_path(self, name: str) -> None:
+        """Remove ``name`` from the ``Local`` namespace, if present."""
+        state_data = self.get_state_data()
+        local = state_data.get("Local")
+        if local is None or name not in local:
+            return
+        local.pop(name, None)
+        self.set_state_data(state_data)
+
     def eval(self, expression: str) -> Any:
         """Evaluate a PowerFx expression with the current state.
 
@@ -504,53 +544,64 @@ class DeclarativeWorkflowState:
             return result
 
         # Pre-process nested custom functions (e.g., Upper(MessageText(...)))
-        # Replace them with their evaluated results before sending to PowerFx
-        formula = self._preprocess_custom_functions(formula)
+        # and run PowerFx. The finally below restores any temporary state
+        # written during preprocessing, regardless of where execution exits.
+        temp_writes: list[tuple[str, Any]] = []
 
-        if Engine is None:
-            raise RuntimeError(
-                f"PowerFx is not available (dotnet runtime not installed). "
-                f"Expression '={formula[:80]}' cannot be evaluated. "
-                f"Install dotnet and the powerfx package for full PowerFx support."
-            )
-
-        symbols = self._to_powerfx_symbols()
-        # Use setlocale(category) query form so we can restore the exact prior value.
-        # getlocale() returns a normalized tuple and is not always a lossless
-        # round-trip for setlocale across platforms/locales.
-        original_numeric_locale = locale.setlocale(locale.LC_NUMERIC)
         try:
-            for locale_candidate in _POWERFX_NUMERIC_LOCALE_CANDIDATES:
-                try:
-                    locale.setlocale(locale.LC_NUMERIC, locale_candidate)
-                    break
-                except locale.Error:
-                    continue
+            formula = self._preprocess_custom_functions(formula, temp_writes)
 
-            engine = Engine()
-            try:
-                from System.Globalization import (  # pyright: ignore[reportMissingImports]
-                    CultureInfo,  # pyright: ignore[reportUnknownVariableType]
+            if Engine is None:
+                raise RuntimeError(
+                    f"PowerFx is not available (dotnet runtime not installed). "
+                    f"Expression '={formula[:80]}' cannot be evaluated. "
+                    f"Install dotnet and the powerfx package for full PowerFx support."
                 )
-            except ImportError:
-                return engine.eval(formula, symbols=symbols, locale=_POWERFX_EVAL_LOCALE)
 
-            original_culture = cast(Any, CultureInfo.CurrentCulture)  # pyright: ignore[reportUnknownMemberType]
+            symbols = self._to_powerfx_symbols()
+            # Use setlocale(category) query form so we can restore the exact prior value.
+            # getlocale() returns a normalized tuple and is not always a lossless
+            # round-trip for setlocale across platforms/locales.
+            original_numeric_locale = locale.setlocale(locale.LC_NUMERIC)
             try:
-                CultureInfo.CurrentCulture = CultureInfo(_POWERFX_EVAL_LOCALE)  # pyright: ignore[reportUnknownMemberType, reportUnknownVariableType]
-                return engine.eval(formula, symbols=symbols, locale=_POWERFX_EVAL_LOCALE)
+                for locale_candidate in _POWERFX_NUMERIC_LOCALE_CANDIDATES:
+                    try:
+                        locale.setlocale(locale.LC_NUMERIC, locale_candidate)
+                        break
+                    except locale.Error:
+                        continue
+
+                engine = Engine()
+                try:
+                    from System.Globalization import (  # pyright: ignore[reportMissingImports]
+                        CultureInfo,  # pyright: ignore[reportUnknownVariableType]
+                    )
+                except ImportError:
+                    return engine.eval(formula, symbols=symbols, locale=_POWERFX_EVAL_LOCALE)
+
+                original_culture = cast(Any, CultureInfo.CurrentCulture)  # pyright: ignore[reportUnknownMemberType]
+                try:
+                    CultureInfo.CurrentCulture = CultureInfo(_POWERFX_EVAL_LOCALE)  # pyright: ignore[reportUnknownMemberType, reportUnknownVariableType]
+                    return engine.eval(formula, symbols=symbols, locale=_POWERFX_EVAL_LOCALE)
+                finally:
+                    CultureInfo.CurrentCulture = original_culture  # pyright: ignore[reportUnknownMemberType]
+            except ValueError as e:
+                error_msg = str(e)
+                # Handle undefined variable errors gracefully by returning None
+                # This matches the behavior of the legacy fallback parser
+                if "isn't recognized" in error_msg or "Name isn't valid" in error_msg:
+                    logger.debug(f"PowerFx: undefined variable in expression '{formula}', returning None")
+                    return None
+                raise
             finally:
-                CultureInfo.CurrentCulture = original_culture  # pyright: ignore[reportUnknownMemberType]
-        except ValueError as e:
-            error_msg = str(e)
-            # Handle undefined variable errors gracefully by returning None
-            # This matches the behavior of the legacy fallback parser
-            if "isn't recognized" in error_msg or "Name isn't valid" in error_msg:
-                logger.debug(f"PowerFx: undefined variable in expression '{formula}', returning None")
-                return None
-            raise
+                locale.setlocale(locale.LC_NUMERIC, original_numeric_locale)
         finally:
-            locale.setlocale(locale.LC_NUMERIC, original_numeric_locale)
+            # Restore each temporary key to its prior value (or remove it).
+            for path, previous in reversed(temp_writes):
+                if previous is self._MISSING:
+                    self._clear_local_path(path.removeprefix("Local."))
+                else:
+                    self.set(path, previous)
 
     def _eval_custom_function(self, formula: str) -> Any | None:
         """Handle custom functions not supported by the Python PowerFx library.
@@ -609,7 +660,7 @@ class DeclarativeWorkflowState:
 
         return None
 
-    def _preprocess_custom_functions(self, formula: str) -> str:
+    def _preprocess_custom_functions(self, formula: str, temp_writes: list[tuple[str, Any]]) -> str:
         """Pre-process custom functions nested inside other PowerFx functions.
 
         Custom functions like MessageText() are not supported by the PowerFx engine.
@@ -624,9 +675,14 @@ class DeclarativeWorkflowState:
 
         Args:
             formula: The PowerFx formula to pre-process
+            temp_writes: Caller-owned list. Each write to a temporary key
+                appends a ``(path, previous_value)`` entry where
+                ``previous_value`` is the value at ``path`` before the write
+                or :attr:`_MISSING` if none. The caller must restore every
+                entry, including when this method raises mid-write.
 
         Returns:
-            The formula with custom function calls replaced by their evaluated results
+            The rewritten formula.
         """
         import re
 
@@ -635,7 +691,6 @@ class DeclarativeWorkflowState:
         # We use 500 to leave room for the rest of the expression around the replaced value.
         MAX_INLINE_LENGTH = 500
 
-        # Counter for generating unique temp variable names
         temp_var_counter = 0
 
         # Custom functions that need pre-processing: (regex pattern, handler)
@@ -691,11 +746,14 @@ class DeclarativeWorkflowState:
                 # Replace in formula
                 if isinstance(replacement, str):
                     if len(replacement) > MAX_INLINE_LENGTH:
-                        # Store long strings in a temp variable to avoid PowerFx expression limit
+                        # Store long results in an underscore-prefixed temp key;
+                        # record the prior value so eval() can restore it.
                         temp_var_name = f"_TempMessageText{temp_var_counter}"
                         temp_var_counter += 1
-                        self.set(f"Local.{temp_var_name}", replacement)
-                        replacement_str = f"Local.{temp_var_name}"
+                        temp_var_path = f"Local.{temp_var_name}"
+                        temp_writes.append((temp_var_path, self.get(temp_var_path, default=self._MISSING)))
+                        self.set(temp_var_path, replacement)
+                        replacement_str = temp_var_path
                         logger.debug(
                             f"Stored long MessageText result ({len(replacement)} chars) "
                             f"in temp variable {temp_var_name}"
@@ -847,11 +905,13 @@ class DeclarativeWorkflowState:
         return value
 
     def interpolate_string(self, text: str) -> str:
-        """Interpolate {Variable.Path} references in a string.
+        """Interpolate ``{Variable.Path}`` references in a string.
 
-        This handles template-style variable substitution like:
-        - "Created ticket #{Local.TicketParameters.TicketId}"
-        - "Routing to {Local.RoutingParameters.TeamName}"
+        Captures brace-delimited tokens whose root segment is an identifier
+        (``[A-Za-z][A-Za-z0-9_]*``) followed by zero or more ``.`` separated
+        dict-key segments. Resolution is delegated to :meth:`get`; unresolved
+        tokens are replaced with the empty string. Tokens that do not look
+        like state paths (e.g. ``{foo-bar}``, ``{Ctrl+C}``) are left literal.
 
         Args:
             text: Text that may contain {Variable.Path} references
@@ -866,10 +926,11 @@ class DeclarativeWorkflowState:
             value = self.get(var_path)
             return str(value) if value is not None else ""
 
-        # Match {Variable.Path} patterns
-        pattern = r"\{([A-Za-z][A-Za-z0-9_.]*)\}"
+        # Root segment must be an identifier; follow-on segments accept any
+        # non-empty dict-key (e.g. ``_id``, ``1``, UUIDs). ``get()`` enforces
+        # per-segment safety on attribute traversal.
+        pattern = r"\{([A-Za-z][A-Za-z0-9_]*(?:\.[^{}\s.]+)*)\}"
 
-        # Replace all matches
         result = text
         for match in re.finditer(pattern, text):
             replacement = replace_var(match)

python/packages/declarative/agent_framework_declarative/_workflows/_executors_mcp.py

@@ -10,17 +10,11 @@ optional conversation history. Supports a human-in-loop approval flow via
 
 Security notes:
 
-- The executor never echoes header VALUES (auth tokens, API keys) into the
-  approval request — only header NAMES are surfaced to the caller. This
-  matches the security posture of :mod:`._executors_http` (which never logs
-  request headers either) and prevents secrets from leaking through workflow
-  events that are typically observable to operators / UIs.
-- ``_MCPToolApprovalState`` snapshots the EVALUATED values for non-secret
-  fields (server URL, tool name, arguments) at approval-request time so that
-  subsequent state mutations cannot make the executor "approve X then call
-  Y". Headers are stored as the raw expression strings (not evaluated values)
-  so secrets are not persisted in the workflow's checkpoint state. They are
-  re-evaluated on resume.
+- Approval requests surface header NAMES only; header values are not echoed,
+  matching the posture of :mod:`._executors_http`.
+- :class:`MCPToolApprovalRequest` carries the values the resume handler will
+  use; header values are re-evaluated on resume to keep secrets out of
+  checkpoint state.
 - Tool outputs flow back into agent conversations through ``conversationId``
   and through Tool-role messages emitted to ``output.messages``. They share
   the same prompt-injection risk surface as ``HttpRequestAction``: workflow
@@ -60,8 +54,6 @@ __all__ = [
 
 logger = logging.getLogger(__name__)
 
-_MCP_APPROVAL_STATE_KEY = "_mcp_tool_approval_state"
-
 
 # ---------------------------------------------------------------------------
 # Request / state types
@@ -72,20 +64,16 @@ _MCP_APPROVAL_STATE_KEY = "_mcp_tool_approval_state"
 class MCPToolApprovalRequest:
     """Approval request emitted before invoking an MCP tool.
 
-    Mirrors :class:`agent_framework_declarative.ToolApprovalRequest` but for
-    MCP-style invocations. Only header NAMES are surfaced — header values are
-    intentionally omitted because they typically carry authentication
-    secrets.
-
     Attributes:
-        request_id: Unique identifier for this approval request. Matches the
-            id workflow event-emitters use.
-        tool_name: Evaluated name of the tool to be invoked.
+        request_id: Identifier matching the framework's pending-request key.
+        tool_name: Evaluated tool name.
         server_url: Evaluated MCP server URL.
-        server_label: Optional human-readable label for diagnostics.
-        arguments: Evaluated arguments to be forwarded to the tool.
-        header_names: Sorted list of outbound header names (no values). Empty
-            when no headers are configured.
+        server_label: Optional human-readable label.
+        arguments: Evaluated tool arguments.
+        header_names: Outbound header names (values withheld).
+        connection_name: Connection identifier the invocation will use.
+        metadata: Internal routing data pinned at approval-request time
+            (e.g. ``conversation_id``) for use by the resume handler.
     """
 
     request_id: str
@@ -94,28 +82,8 @@ class MCPToolApprovalRequest:
     server_label: str | None
     arguments: dict[str, Any]
     header_names: list[str] = field(default_factory=lambda: [])
-
-
-@dataclass
-class _MCPToolApprovalState:
-    """Internal state saved during the approval yield for resumption.
-
-    Stores **evaluated** values for non-secret fields to prevent
-    "approve X / execute Y" attacks. Stores the raw expression string for
-    ``headers`` so that secret values are NOT persisted in checkpoint state;
-    the expressions are re-evaluated against current state on resume.
-    """
-
-    server_url: str
-    tool_name: str
-    server_label: str | None
-    arguments: dict[str, Any]
-    connection_name: str | None
-    headers_def: Any
-    auto_send: bool
-    conversation_id_expr: str | None
-    output_messages_path: str | None
-    output_result_path: str | None
+    connection_name: str | None = None
+    metadata: dict[str, Any] = field(default_factory=lambda: {})
 
 
 # ---------------------------------------------------------------------------
@@ -123,21 +91,15 @@ class _MCPToolApprovalState:
 # ---------------------------------------------------------------------------
 
 
-def _get_messages_path(state: DeclarativeWorkflowState, conversation_id_expr: str | None) -> str | None:
-    """Return the configured conversation messages path, if any.
-
-    Returns ``System.conversations.{evaluated_id}.messages`` when a
-    ``conversation_id_expr`` is configured and evaluates to a non-empty value.
-    Returns ``None`` when no conversation id expression is configured or when
-    the expression evaluates to ``None`` or an empty string (mirrors .NET
-    ``GetConversationId`` behaviour).
-    """
-    if not conversation_id_expr:
+def _evaluate_conversation_id(state: DeclarativeWorkflowState, conversation_id_expr: Any) -> str | None:
+    """Return the evaluated ``conversationId`` string, or None when empty/unset."""
+    if not isinstance(conversation_id_expr, str) or not conversation_id_expr:
         return None
     evaluated = state.eval_if_expression(conversation_id_expr)
-    if evaluated is None or (isinstance(evaluated, str) and not evaluated):
+    if evaluated is None:
         return None
-    return f"System.conversations.{evaluated}.messages"
+    text = str(evaluated)
+    return text or None
 
 
 def _get_output_path(action_def: Mapping[str, Any], key: str) -> str | None:
@@ -260,20 +222,7 @@ class InvokeMcpToolActionExecutor(DeclarativeActionExecutor):
 
         if require_approval:
             request_id = str(uuid.uuid4())
-            approval_state = _MCPToolApprovalState(
-                server_url=server_url,
-                tool_name=tool_name,
-                server_label=server_label,
-                arguments=arguments,
-                connection_name=connection_name,
-                headers_def=self._action_def.get("headers"),
-                auto_send=auto_send,
-                conversation_id_expr=conversation_id_expr if isinstance(conversation_id_expr, str) else None,
-                output_messages_path=output_messages_path,
-                output_result_path=output_result_path,
-            )
-            ctx.state.set(self._approval_key(), approval_state)
-
+            conversation_id = _evaluate_conversation_id(state, conversation_id_expr)
             request = MCPToolApprovalRequest(
                 request_id=request_id,
                 tool_name=tool_name,
@@ -281,6 +230,8 @@ class InvokeMcpToolActionExecutor(DeclarativeActionExecutor):
                 server_label=server_label,
                 arguments=arguments,
                 header_names=sorted(headers.keys()),
+                connection_name=connection_name,
+                metadata={"conversation_id": conversation_id},
             )
             logger.info(
                 "%s: requesting approval for MCP tool '%s' on '%s'",
@@ -289,7 +240,6 @@ class InvokeMcpToolActionExecutor(DeclarativeActionExecutor):
                 server_url,
             )
             await ctx.request_info(request, ToolApprovalResponse, request_id=request_id)
-            # Workflow yields here — resume in handle_approval_response.
             return
 
         # No approval required - invoke directly.
@@ -307,7 +257,7 @@ class InvokeMcpToolActionExecutor(DeclarativeActionExecutor):
             state=state,
             result=result,
             auto_send=auto_send,
-            conversation_id_expr=conversation_id_expr if isinstance(conversation_id_expr, str) else None,
+            conversation_id=_evaluate_conversation_id(state, conversation_id_expr),
             output_messages_path=output_messages_path,
             output_result_path=output_result_path,
         )
@@ -322,54 +272,46 @@ class InvokeMcpToolActionExecutor(DeclarativeActionExecutor):
         response: ToolApprovalResponse,
         ctx: WorkflowContext[ActionComplete, str],
     ) -> None:
-        """Resume after the workflow yielded for an approval request."""
+        """Resume the invocation using the values pinned on ``original_request``."""
         state = self._get_state(ctx.state)
-        approval_key = self._approval_key()
 
-        try:
-            approval_state: _MCPToolApprovalState = ctx.state.get(approval_key)
-        except KeyError:
-            logger.error("%s: approval state missing for executor '%s'", self.__class__.__name__, self.id)
-            await ctx.send_message(ActionComplete())
-            return
-        try:
-            ctx.state.delete(approval_key)
-        except KeyError:
-            logger.warning("%s: approval state already deleted for '%s'", self.__class__.__name__, self.id)
+        tool_name = original_request.tool_name
+        metadata: dict[str, Any] = getattr(original_request, "metadata", None) or {}
+        raw_conversation_id = metadata.get("conversation_id")
+        conversation_id = raw_conversation_id if isinstance(raw_conversation_id, str) and raw_conversation_id else None
+
+        auto_send = self._get_auto_send(state)
+        output_messages_path = _get_output_path(self._action_def, "messages")
+        output_result_path = _get_output_path(self._action_def, "result")
 
         if not response.approved:
             logger.info(
                 "%s: MCP tool '%s' rejected: %s",
                 self.__class__.__name__,
-                approval_state.tool_name,
+                tool_name,
                 response.reason,
             )
-            self._assign_error(
-                state, approval_state.output_result_path, "MCP tool invocation was not approved by user."
-            )
+            self._assign_error(state, output_result_path, "MCP tool invocation was not approved by user.")
             await ctx.send_message(ActionComplete())
             return
 
-        # Approved — re-evaluate headers (not stored at approval time for security).
-        headers = self._evaluate_headers(state, approval_state.headers_def)
-
         invocation = MCPToolInvocation(
-            server_url=approval_state.server_url,
-            tool_name=approval_state.tool_name,
-            server_label=approval_state.server_label,
-            arguments=approval_state.arguments,
-            headers=headers,
-            connection_name=approval_state.connection_name,
+            server_url=original_request.server_url,
+            tool_name=tool_name,
+            server_label=original_request.server_label,
+            arguments=original_request.arguments,
+            headers=self._evaluate_headers(state, self._action_def.get("headers")),
+            connection_name=getattr(original_request, "connection_name", None),
         )
         result = await self._invoke_with_narrow_catch(invocation)
         await self._process_result(
             ctx=ctx,
             state=state,
             result=result,
-            auto_send=approval_state.auto_send,
-            conversation_id_expr=approval_state.conversation_id_expr,
-            output_messages_path=approval_state.output_messages_path,
-            output_result_path=approval_state.output_result_path,
+            auto_send=auto_send,
+            conversation_id=conversation_id,
+            output_messages_path=output_messages_path,
+            output_result_path=output_result_path,
         )
         await ctx.send_message(ActionComplete())
 
@@ -528,7 +470,7 @@ class InvokeMcpToolActionExecutor(DeclarativeActionExecutor):
         state: DeclarativeWorkflowState,
         result: MCPToolResult,
         auto_send: bool,
-        conversation_id_expr: str | None,
+        conversation_id: str | None,
         output_messages_path: str | None,
         output_result_path: str | None,
     ) -> None:
@@ -557,14 +499,10 @@ class InvokeMcpToolActionExecutor(DeclarativeActionExecutor):
         if auto_send and parsed_results:
             await ctx.yield_output(_format_outputs_for_send(parsed_results))
 
-        if conversation_id_expr:
-            messages_path = _get_messages_path(state, conversation_id_expr)
-            if messages_path is not None:
-                # Mirrors .NET: conversation gets ASSISTANT-role message with
-                # the same outputs (so chat history reads it as the agent's
-                # contribution).
-                assistant_message = Message(role="assistant", contents=list(result.outputs))
-                state.append(messages_path, assistant_message)
+        if conversation_id:
+            messages_path = f"System.conversations.{conversation_id}.messages"
+            assistant_message = Message(role="assistant", contents=list(result.outputs))
+            state.append(messages_path, assistant_message)
 
     @staticmethod
     def _assign_error(
@@ -577,9 +515,6 @@ class InvokeMcpToolActionExecutor(DeclarativeActionExecutor):
             return
         state.set(output_result_path, f"Error: {error_message}")
 
-    def _approval_key(self) -> str:
-        return f"{_MCP_APPROVAL_STATE_KEY}_{self.id}"
-
 
 def _parse_outputs(outputs: list[Content]) -> list[Any]:
     """Parse :class:`Content` outputs into Python values for ``output.result``.

python/packages/declarative/agent_framework_declarative/_workflows/_executors_tools.py

@@ -41,10 +41,6 @@ logger = logging.getLogger(__name__)
 # at runtime are discoverable by both agent-based and function-based tool executors.
 FUNCTION_TOOL_REGISTRY_KEY = TOOL_REGISTRY_KEY
 
-# State key prefix for storing approval state during yield/resume.
-# The executor's ID is appended to create a per-executor key.
-TOOL_APPROVAL_STATE_KEY = "_tool_approval_state"
-
 
 # ============================================================================
 # Request/Response Types for Approval Flow
@@ -87,26 +83,6 @@ class ToolApprovalResponse:
     reason: str | None = None
 
 
-# ============================================================================
-# State Types for Approval Flow
-# ============================================================================
-
-
-@dataclass
-class ToolApprovalState:
-    """State saved during approval yield for resumption.
-
-    Stored in State under a per-executor key when requireApproval=true.
-    Retrieved by handle_approval_response() to continue execution.
-    """
-
-    function_name: str
-    arguments: dict[str, Any]
-    output_messages_var: str | None
-    output_result_var: str | None
-    auto_send: bool
-
-
 # ============================================================================
 # Result Types
 # ============================================================================
@@ -501,25 +477,16 @@ class BaseToolExecutor(DeclarativeActionExecutor):
         require_approval = self._action_def.get("requireApproval", False)
 
         if require_approval:
-            # Save state for resumption (keyed by executor ID to avoid collisions)
-            approval_state = ToolApprovalState(
-                function_name=function_name,
-                arguments=arguments,
-                output_messages_var=messages_var,
-                output_result_var=result_var,
-                auto_send=auto_send,
-            )
-            approval_key = f"{TOOL_APPROVAL_STATE_KEY}_{self.id}"
-            ctx.state.set(approval_key, approval_state)
-
-            # Emit approval request - workflow yields here
+            # Emit approval request - the request payload is the source of
+            # truth for resumed invocation; no side-channel state is written.
+            request_id = str(uuid.uuid4())
             request = ToolApprovalRequest(
-                request_id=str(uuid.uuid4()),
+                request_id=request_id,
                 function_name=function_name,
                 arguments=arguments,
             )
             logger.info(f"{self.__class__.__name__}: requesting approval for '{function_name}'")
-            await ctx.request_info(request, ToolApprovalResponse)
+            await ctx.request_info(request, ToolApprovalResponse, request_id=request_id)
             # Workflow yields - will resume in handle_approval_response
             return
 
@@ -545,36 +512,16 @@ class BaseToolExecutor(DeclarativeActionExecutor):
     ) -> None:
         """Handle response to a ToolApprovalRequest.
 
-        Called when the workflow resumes after yielding for approval.
-        Either executes the tool (if approved) or stores rejection status.
+        Resumes after the workflow yielded for approval. The invocation
+        ``function_name`` and ``arguments`` are sourced from
+        ``original_request`` (the payload the reviewer approved); output
+        configuration is re-derived from the executor's action definition.
         """
         state = self._get_state(ctx.state)
-        approval_key = f"{TOOL_APPROVAL_STATE_KEY}_{self.id}"
 
-        # Retrieve saved invocation state
-        try:
-            approval_state: ToolApprovalState = ctx.state.get(approval_key)
-        except KeyError:
-            error_msg = "Approval state not found, cannot resume tool invocation"
-            logger.error(f"{self.__class__.__name__}: {error_msg}")
-            # Try to store error - get output config from action def as fallback
-            _, result_var, _ = self._get_output_config()
-            if result_var and state:
-                state.set(_normalize_variable_path(result_var), {"error": error_msg})
-            await ctx.send_message(ActionComplete())
-            return
-
-        # Clean up approval state
-        try:
-            ctx.state.delete(approval_key)
-        except KeyError:
-            logger.warning(f"{self.__class__.__name__}: approval state already deleted")
-
-        function_name = approval_state.function_name
-        arguments = approval_state.arguments
-        messages_var = approval_state.output_messages_var
-        result_var = approval_state.output_result_var
-        auto_send = approval_state.auto_send
+        function_name = original_request.function_name
+        arguments = original_request.arguments
+        messages_var, result_var, auto_send = self._get_output_config()
 
         # Check if approved
         if not response.approved:

python/packages/declarative/tests/test_declarative_approval_binding.py

@@ -0,0 +1,528 @@
+# Copyright (c) Microsoft. All rights reserved.
+# pyright: reportUnknownParameterType=false, reportUnknownArgumentType=false
+# pyright: reportMissingParameterType=false, reportUnknownMemberType=false
+# pyright: reportPrivateUsage=false, reportUnknownVariableType=false
+# pyright: reportGeneralTypeIssues=false
+
+"""Regression tests pinning the approval-flow binding contract.
+
+The resumed invocation MUST come from the framework-delivered
+``original_request`` payload (the data the reviewer approved) for both
+``InvokeFunctionTool`` and ``InvokeMcpTool``. These tests verify that:
+
+* Invocation parameters come from ``original_request``, not from any prior
+  side-channel state.
+* Concurrent pending approvals on the same executor do not swap.
+* Pre-existing state at old approval keys is ignored entirely.
+* Resume works on a freshly constructed executor (checkpoint-restore
+  simulation), without any prior ``ctx.state`` write.
+* For MCP, ``connection_name`` is sourced from the approval payload and
+  ``headers`` are re-evaluated from the action definition on resume.
+"""
+
+import sys
+from dataclasses import dataclass
+from typing import Any
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+
+try:
+    import powerfx  # noqa: F401
+
+    _powerfx_available = True
+except (ImportError, RuntimeError):
+    _powerfx_available = False
+
+pytestmark = pytest.mark.skipif(
+    not _powerfx_available or sys.version_info >= (3, 14),
+    reason="PowerFx engine not available (requires dotnet runtime)",
+)
+
+from agent_framework import Content  # noqa: E402
+
+from agent_framework_declarative._workflows import (  # noqa: E402
+    DECLARATIVE_STATE_KEY,
+    ActionComplete,
+    InvokeFunctionToolExecutor,
+    MCPToolApprovalRequest,
+    MCPToolHandler,
+    MCPToolInvocation,
+    MCPToolResult,
+    ToolApprovalRequest,
+    ToolApprovalResponse,
+)
+from agent_framework_declarative._workflows._declarative_base import DeclarativeWorkflowState  # noqa: E402
+from agent_framework_declarative._workflows._executors_mcp import (  # noqa: E402
+    InvokeMcpToolActionExecutor,
+)
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture
+def mock_state() -> MagicMock:
+    """In-memory mock of the underlying State."""
+    state = MagicMock()
+    state._data = {}
+
+    def _get(key: str, default: Any = None) -> Any:
+        return state._data.get(key, default)
+
+    def _set(key: str, value: Any) -> None:
+        state._data[key] = value
+
+    def _has(key: str) -> bool:
+        return key in state._data
+
+    def _delete(key: str) -> None:
+        state._data.pop(key, None)
+
+    state.get = MagicMock(side_effect=_get)
+    state.set = MagicMock(side_effect=_set)
+    state.has = MagicMock(side_effect=_has)
+    state.delete = MagicMock(side_effect=_delete)
+    return state
+
+
+@pytest.fixture
+def mock_context(mock_state: MagicMock) -> MagicMock:
+    ctx = MagicMock()
+    ctx.state = mock_state
+    ctx.send_message = AsyncMock()
+    ctx.yield_output = AsyncMock()
+    ctx.request_info = AsyncMock()
+    return ctx
+
+
+def _seed_state(mock_state: MagicMock) -> None:
+    mock_state._data[DECLARATIVE_STATE_KEY] = {
+        "Inputs": {},
+        "Outputs": {},
+        "Local": {},
+        "Custom": {},
+        "System": {
+            "ConversationId": "00000000-0000-0000-0000-000000000000",
+            "LastMessage": {"Text": "", "Id": ""},
+            "LastMessageText": "",
+            "LastMessageId": "",
+        },
+        "Agent": {},
+        "Conversation": {"messages": [], "history": []},
+    }
+
+
+class _RecordingMcpHandler(MCPToolHandler):
+    def __init__(self, result: MCPToolResult | None = None) -> None:
+        self.result = result or MCPToolResult(outputs=[Content.from_text("ok")])
+        self.invocations: list[MCPToolInvocation] = []
+
+    @property
+    def call_count(self) -> int:
+        return len(self.invocations)
+
+    @property
+    def last(self) -> MCPToolInvocation | None:
+        return self.invocations[-1] if self.invocations else None
+
+    async def invoke_tool(self, invocation: MCPToolInvocation) -> MCPToolResult:
+        self.invocations.append(invocation)
+        return self.result
+
+
+# ---------------------------------------------------------------------------
+# InvokeFunctionTool: approval-binding regression
+# ---------------------------------------------------------------------------
+
+
+class TestFunctionToolApprovalBinding:
+    def _action(self, *, fn_name: str = "my_tool") -> dict[str, Any]:
+        return {
+            "kind": "InvokeFunctionTool",
+            "id": "fn_action",
+            "functionName": fn_name,
+            "requireApproval": True,
+            "output": {"result": "Local.result"},
+        }
+
+    @pytest.mark.asyncio
+    async def test_request_id_matches_framework_pending_key(self, mock_state, mock_context) -> None:
+        """The id on the emitted ToolApprovalRequest must match the framework's pending-request key."""
+        from agent_framework_declarative._workflows._declarative_base import ActionTrigger
+
+        _seed_state(mock_state)
+
+        def my_tool(x: int) -> int:
+            return x
+
+        executor = InvokeFunctionToolExecutor(self._action(), tools={"my_tool": my_tool})
+        await executor.handle_action(ActionTrigger(), mock_context)
+
+        mock_context.request_info.assert_called_once()
+        emitted_request = mock_context.request_info.call_args[0][0]
+        framework_request_id = mock_context.request_info.call_args.kwargs["request_id"]
+        assert isinstance(emitted_request, ToolApprovalRequest)
+        assert emitted_request.request_id == framework_request_id
+
+    @pytest.mark.asyncio
+    async def test_resume_uses_request_payload_arguments(self, mock_state, mock_context) -> None:
+        _seed_state(mock_state)
+        call_log: list[int] = []
+
+        def my_tool(x: int) -> int:
+            call_log.append(x)
+            return x
+
+        executor = InvokeFunctionToolExecutor(self._action(), tools={"my_tool": my_tool})
+
+        request = ToolApprovalRequest(request_id="r-1", function_name="my_tool", arguments={"x": 1})
+        await executor.handle_approval_response(request, ToolApprovalResponse(approved=True), mock_context)
+
+        assert call_log == [1]
+
+    @pytest.mark.asyncio
+    async def test_concurrent_pending_approvals_do_not_swap(self, mock_state, mock_context) -> None:
+        """Two pending approvals, responses delivered out of order — each invocation uses its own payload."""
+        _seed_state(mock_state)
+        call_log: list[int] = []
+
+        def my_tool(x: int) -> int:
+            call_log.append(x)
+            return x
+
+        executor = InvokeFunctionToolExecutor(self._action(), tools={"my_tool": my_tool})
+
+        request_a = ToolApprovalRequest(request_id="r-A", function_name="my_tool", arguments={"x": 1})
+        request_b = ToolApprovalRequest(request_id="r-B", function_name="my_tool", arguments={"x": 999})
+
+        # Deliver response for B first, then for A. Each invocation must use its own payload.
+        await executor.handle_approval_response(request_b, ToolApprovalResponse(approved=True), mock_context)
+        await executor.handle_approval_response(request_a, ToolApprovalResponse(approved=True), mock_context)
+
+        assert call_log == [999, 1]
+
+    @pytest.mark.asyncio
+    async def test_resume_ignores_stale_state_at_old_approval_key(self, mock_state, mock_context) -> None:
+        """Pre-existing state at the OLD approval key is ignored — payload wins."""
+        _seed_state(mock_state)
+        call_log: list[int] = []
+
+        def my_tool(x: int) -> int:
+            call_log.append(x)
+            return x
+
+        executor = InvokeFunctionToolExecutor(self._action(), tools={"my_tool": my_tool})
+
+        # Poison the old key shape (no longer read by the executor).
+        mock_state._data["_tool_approval_state_fn_action"] = {"function_name": "other", "arguments": {"x": 999}}
+
+        request = ToolApprovalRequest(request_id="r-3", function_name="my_tool", arguments={"x": 7})
+        await executor.handle_approval_response(request, ToolApprovalResponse(approved=True), mock_context)
+
+        assert call_log == [7]
+        # The poison was never read or deleted by the executor.
+        assert "_tool_approval_state_fn_action" in mock_state._data
+
+    @pytest.mark.asyncio
+    async def test_fresh_executor_resume_works(self, mock_state, mock_context) -> None:
+        """Simulates checkpoint restore: a brand-new executor instance handles the approval response."""
+        _seed_state(mock_state)
+        call_log: list[int] = []
+
+        def my_tool(x: int) -> int:
+            call_log.append(x)
+            return x
+
+        # Pretend the executor that emitted the request is gone; a fresh one handles the response.
+        fresh = InvokeFunctionToolExecutor(self._action(), tools={"my_tool": my_tool})
+
+        request = ToolApprovalRequest(request_id="r-4", function_name="my_tool", arguments={"x": 42})
+        await fresh.handle_approval_response(request, ToolApprovalResponse(approved=True), mock_context)
+
+        assert call_log == [42]
+        mock_context.send_message.assert_called_once()
+        sent = mock_context.send_message.call_args[0][0]
+        assert isinstance(sent, ActionComplete)
+
+    @pytest.mark.asyncio
+    async def test_rejection_uses_request_payload_function_name(self, mock_state, mock_context) -> None:
+        _seed_state(mock_state)
+
+        def my_tool(x: int) -> int:
+            raise AssertionError("should not be called when rejected")
+
+        executor = InvokeFunctionToolExecutor(self._action(), tools={"my_tool": my_tool})
+
+        request = ToolApprovalRequest(request_id="r-5", function_name="my_tool", arguments={"x": 3})
+        await executor.handle_approval_response(
+            request, ToolApprovalResponse(approved=False, reason="not authorized"), mock_context
+        )
+
+        # The rejection message references the function name from the request payload.
+        local = mock_state._data[DECLARATIVE_STATE_KEY]["Local"]
+        assert local["result"]["rejected"] is True
+        assert local["result"]["reason"] == "not authorized"
+
+
+# ---------------------------------------------------------------------------
+# InvokeMcpTool: approval-binding regression
+# ---------------------------------------------------------------------------
+
+
+class TestMcpToolApprovalBinding:
+    def _action(self, *, headers: dict[str, Any] | None = None) -> dict[str, Any]:
+        action: dict[str, Any] = {
+            "kind": "InvokeMcpTool",
+            "id": "mcp_action",
+            "serverUrl": "https://mcp.example/api",
+            "toolName": "search",
+            "requireApproval": True,
+            "output": {"result": "Local.Result"},
+        }
+        if headers is not None:
+            action["headers"] = headers
+        return action
+
+    @pytest.mark.asyncio
+    async def test_request_id_matches_framework_pending_key(self, mock_state, mock_context) -> None:
+        """The id on the emitted MCPToolApprovalRequest must match the framework's pending-request key."""
+        from agent_framework_declarative._workflows._declarative_base import ActionTrigger
+
+        _seed_state(mock_state)
+        executor = InvokeMcpToolActionExecutor(self._action(), mcp_tool_handler=_RecordingMcpHandler())
+        await executor.handle_action(ActionTrigger(), mock_context)
+
+        mock_context.request_info.assert_called_once()
+        emitted_request = mock_context.request_info.call_args[0][0]
+        framework_request_id = mock_context.request_info.call_args.kwargs["request_id"]
+        assert isinstance(emitted_request, MCPToolApprovalRequest)
+        assert emitted_request.request_id == framework_request_id
+
+    @pytest.mark.asyncio
+    async def test_resume_uses_request_payload_fields(self, mock_state, mock_context) -> None:
+        _seed_state(mock_state)
+        handler = _RecordingMcpHandler()
+        executor = InvokeMcpToolActionExecutor(self._action(), mcp_tool_handler=handler)
+
+        request = MCPToolApprovalRequest(
+            request_id="r-1",
+            tool_name="search",
+            server_url="https://mcp.example/api",
+            server_label="prod",
+            arguments={"q": "x"},
+            connection_name="conn-A",
+        )
+        await executor.handle_approval_response(request, ToolApprovalResponse(approved=True), mock_context)
+
+        assert handler.call_count == 1
+        inv = handler.last
+        assert inv is not None
+        assert inv.tool_name == "search"
+        assert inv.server_url == "https://mcp.example/api"
+        assert inv.server_label == "prod"
+        assert inv.arguments == {"q": "x"}
+        assert inv.connection_name == "conn-A"
+
+    @pytest.mark.asyncio
+    async def test_concurrent_pending_mcp_approvals_do_not_swap(self, mock_state, mock_context) -> None:
+        _seed_state(mock_state)
+        handler = _RecordingMcpHandler()
+        executor = InvokeMcpToolActionExecutor(self._action(), mcp_tool_handler=handler)
+
+        request_a = MCPToolApprovalRequest(
+            request_id="r-A",
+            tool_name="search",
+            server_url="https://mcp.example/api",
+            server_label=None,
+            arguments={"q": "alpha"},
+            connection_name="conn-A",
+        )
+        request_b = MCPToolApprovalRequest(
+            request_id="r-B",
+            tool_name="search",
+            server_url="https://mcp.example/api",
+            server_label=None,
+            arguments={"q": "beta"},
+            connection_name="conn-B",
+        )
+
+        await executor.handle_approval_response(request_b, ToolApprovalResponse(approved=True), mock_context)
+        await executor.handle_approval_response(request_a, ToolApprovalResponse(approved=True), mock_context)
+
+        assert handler.call_count == 2
+        assert handler.invocations[0].arguments == {"q": "beta"}
+        assert handler.invocations[0].connection_name == "conn-B"
+        assert handler.invocations[1].arguments == {"q": "alpha"}
+        assert handler.invocations[1].connection_name == "conn-A"
+
+    @pytest.mark.asyncio
+    async def test_headers_reevaluated_from_action_def_on_resume(self, mock_state, mock_context) -> None:
+        """Headers come from the action definition (re-evaluated) so secrets are not in the payload."""
+        _seed_state(mock_state)
+        handler = _RecordingMcpHandler()
+        executor = InvokeMcpToolActionExecutor(
+            self._action(headers={"Authorization": "Bearer tk"}),
+            mcp_tool_handler=handler,
+        )
+
+        request = MCPToolApprovalRequest(
+            request_id="r-1",
+            tool_name="search",
+            server_url="https://mcp.example/api",
+            server_label=None,
+            arguments={"q": "x"},
+            connection_name=None,
+        )
+        await executor.handle_approval_response(request, ToolApprovalResponse(approved=True), mock_context)
+
+        assert handler.last is not None
+        assert handler.last.headers == {"Authorization": "Bearer tk"}
+
+    @pytest.mark.asyncio
+    async def test_mcp_resume_ignores_stale_state_at_old_approval_key(self, mock_state, mock_context) -> None:
+        _seed_state(mock_state)
+        handler = _RecordingMcpHandler()
+        executor = InvokeMcpToolActionExecutor(self._action(), mcp_tool_handler=handler)
+
+        mock_state._data["_mcp_tool_approval_state_mcp_action"] = {"poison": True}
+
+        request = MCPToolApprovalRequest(
+            request_id="r-1",
+            tool_name="search",
+            server_url="https://mcp.example/api",
+            server_label=None,
+            arguments={"q": "real"},
+            connection_name=None,
+        )
+        await executor.handle_approval_response(request, ToolApprovalResponse(approved=True), mock_context)
+
+        assert handler.call_count == 1
+        assert handler.last is not None
+        assert handler.last.arguments == {"q": "real"}
+        # The poison was never read or deleted by the executor.
+        assert "_mcp_tool_approval_state_mcp_action" in mock_state._data
+
+    @pytest.mark.asyncio
+    async def test_fresh_mcp_executor_resume_works(self, mock_state, mock_context) -> None:
+        """Checkpoint-restore simulation: fresh executor handles the response."""
+        _seed_state(mock_state)
+        handler = _RecordingMcpHandler()
+        fresh = InvokeMcpToolActionExecutor(self._action(), mcp_tool_handler=handler)
+
+        request = MCPToolApprovalRequest(
+            request_id="r-1",
+            tool_name="search",
+            server_url="https://mcp.example/api",
+            server_label=None,
+            arguments={"q": "fresh"},
+            connection_name=None,
+        )
+        await fresh.handle_approval_response(request, ToolApprovalResponse(approved=True), mock_context)
+
+        assert handler.call_count == 1
+        assert handler.last is not None
+        assert handler.last.arguments == {"q": "fresh"}
+
+    @pytest.mark.asyncio
+    async def test_request_payload_carries_connection_name(self, mock_state, mock_context) -> None:
+        """When emitting the approval request, connection_name flows into MCPToolApprovalRequest."""
+        from agent_framework_declarative._workflows._declarative_base import ActionTrigger
+
+        _seed_state(mock_state)
+        action = self._action()
+        action["connection"] = {"name": "conn-from-action"}
+        executor = InvokeMcpToolActionExecutor(action, mcp_tool_handler=_RecordingMcpHandler())
+
+        await executor.handle_action(ActionTrigger(), mock_context)
+
+        mock_context.request_info.assert_called_once()
+        request = mock_context.request_info.call_args[0][0]
+        assert isinstance(request, MCPToolApprovalRequest)
+        assert request.connection_name == "conn-from-action"
+
+    @pytest.mark.asyncio
+    async def test_request_payload_pins_conversation_id(self, mock_state, mock_context) -> None:
+        """Evaluated ``conversationId`` is pinned in ``metadata`` at request-emit time."""
+        from agent_framework_declarative._workflows._declarative_base import ActionTrigger
+
+        _seed_state(mock_state)
+        state = DeclarativeWorkflowState(mock_state)
+        state.set("Local.targetConversation", "conv-original")
+        action = self._action()
+        action["conversationId"] = "=Local.targetConversation"
+        executor = InvokeMcpToolActionExecutor(action, mcp_tool_handler=_RecordingMcpHandler())
+
+        await executor.handle_action(ActionTrigger(), mock_context)
+
+        mock_context.request_info.assert_called_once()
+        request = mock_context.request_info.call_args[0][0]
+        assert isinstance(request, MCPToolApprovalRequest)
+        assert request.metadata.get("conversation_id") == "conv-original"
+
+    @pytest.mark.asyncio
+    async def test_resume_routes_output_to_pinned_conversation_not_mutated_state(
+        self, mock_state, mock_context
+    ) -> None:
+        """Output appends to the conversation pinned on ``original_request``, not the
+        current state evaluation."""
+        _seed_state(mock_state)
+        state = DeclarativeWorkflowState(mock_state)
+        state.set("System.conversations.conv-original.messages", [])
+        state.set("System.conversations.conv-mutated.messages", [])
+        state.set("Local.targetConversation", "conv-mutated")
+
+        handler = _RecordingMcpHandler(MCPToolResult(outputs=[Content.from_text("approved-output")]))
+        action = self._action()
+        action["conversationId"] = "=Local.targetConversation"
+        executor = InvokeMcpToolActionExecutor(action, mcp_tool_handler=handler)
+
+        original_request = MCPToolApprovalRequest(
+            request_id="r-1",
+            tool_name="search",
+            server_url="https://mcp.example/api",
+            server_label=None,
+            arguments={"q": "x"},
+            connection_name=None,
+            metadata={"conversation_id": "conv-original"},
+        )
+        await executor.handle_approval_response(original_request, ToolApprovalResponse(approved=True), mock_context)
+
+        assert len(state.get("System.conversations.conv-original.messages") or []) == 1
+        assert state.get("System.conversations.conv-mutated.messages") == []
+
+    @pytest.mark.asyncio
+    async def test_resume_handles_legacy_request_without_new_fields(self, mock_state, mock_context) -> None:
+        """Resume tolerates payloads lacking ``connection_name`` / ``metadata`` (legacy pickle shape)."""
+
+        @dataclass
+        class _LegacyMCPApprovalRequest:
+            request_id: str
+            tool_name: str
+            server_url: str
+            server_label: str | None
+            arguments: dict[str, Any]
+            header_names: list[str]
+
+        _seed_state(mock_state)
+        handler = _RecordingMcpHandler()
+        executor = InvokeMcpToolActionExecutor(self._action(), mcp_tool_handler=handler)
+
+        legacy_request = _LegacyMCPApprovalRequest(
+            request_id="r-1",
+            tool_name="search",
+            server_url="https://mcp.example/api",
+            server_label=None,
+            arguments={"q": "x"},
+            header_names=[],
+        )
+        await executor.handle_approval_response(
+            legacy_request,  # type: ignore[arg-type]
+            ToolApprovalResponse(approved=True),
+            mock_context,
+        )
+
+        assert handler.call_count == 1
+        assert handler.last is not None
+        assert handler.last.connection_name is None

python/packages/declarative/tests/test_declarative_state_path_safety.py

@@ -0,0 +1,364 @@
+# Copyright (c) Microsoft. All rights reserved.
+# pyright: reportUnknownParameterType=false, reportUnknownArgumentType=false
+# pyright: reportMissingParameterType=false, reportUnknownMemberType=false
+# pyright: reportPrivateUsage=false, reportUnknownVariableType=false
+# pyright: reportGeneralTypeIssues=false
+
+"""Path-segment validation tests for DeclarativeWorkflowState.
+
+Path segments handed to ``get``/``set``/``append`` and ``{Variable.Path}``
+placeholders in ``interpolate_string`` are subject to three distinct rules
+that this module pins:
+
+- **Empty segments** (e.g. ``""``, ``"Local."``, ``"Local..foo"``) are rejected
+  by all of ``get``/``set``/``append`` and ``interpolate_string``. ``get`` and
+  ``interpolate_string`` return their default / leave the placeholder literal;
+  ``set`` and ``append`` raise ``ValueError``.
+- **Object-attribute segments** — segments that ``get`` would resolve via
+  ``getattr`` because the parent is a non-dict object — must match the safe
+  identifier shape ``[A-Za-z][A-Za-z0-9_]*``. Other shapes are rejected with a
+  warning log and the default is returned.
+- **Dict-keyed segments** — segments that resolve via dict lookup because the
+  parent is a ``dict`` — may use arbitrary non-empty string keys (e.g. UUIDs
+  or hyphenated identifiers like ``System.conversations.<uuid>.messages``).
+"""
+
+import logging
+from dataclasses import dataclass
+from typing import Any
+from unittest.mock import MagicMock
+
+import pytest
+
+from agent_framework_declarative._workflows import DeclarativeWorkflowState
+
+try:
+    import powerfx  # noqa: F401
+
+    _powerfx_available = True
+except (ImportError, RuntimeError):
+    _powerfx_available = False
+
+_requires_powerfx = pytest.mark.skipif(not _powerfx_available, reason="PowerFx engine not available")
+
+
+@pytest.fixture
+def mock_state() -> MagicMock:
+    """In-memory mock for the underlying State."""
+    ms = MagicMock()
+    ms._data = {}
+
+    def get(key: str, default: Any = None) -> Any:
+        return ms._data.get(key, default)
+
+    def set_(key: str, value: Any) -> None:
+        ms._data[key] = value
+
+    def has(key: str) -> bool:
+        return key in ms._data
+
+    def delete(key: str) -> None:
+        ms._data.pop(key, None)
+
+    ms.get = MagicMock(side_effect=get)
+    ms.set = MagicMock(side_effect=set_)
+    ms.has = MagicMock(side_effect=has)
+    ms.delete = MagicMock(side_effect=delete)
+    return ms
+
+
+@pytest.fixture
+def state(mock_state: MagicMock) -> DeclarativeWorkflowState:
+    s = DeclarativeWorkflowState(mock_state)
+    s.initialize()
+    return s
+
+
+@dataclass
+class _PlainObj:
+    """Non-dict object so ``get`` falls through to attribute access."""
+
+    text: str = "hi"
+
+
+# ---------------------------------------------------------------------------
+# get(): invalid paths return default
+# ---------------------------------------------------------------------------
+
+
+class TestGetRejectsInvalidPaths:
+    def test_rejects_dunder_segment_via_attribute_access(self, state: DeclarativeWorkflowState) -> None:
+        state.set("Local.obj", _PlainObj())
+        assert state.get("Local.obj.__class__") is None
+        assert state.get("Local.obj.__class__", default="DEF") == "DEF"
+
+    def test_rejects_full_env_exfil_chain(self, state: DeclarativeWorkflowState, monkeypatch) -> None:
+        sentinel = "agent-framework-path-safety-sentinel"
+        monkeypatch.setenv("AF_PATH_SAFETY_SENTINEL", sentinel)
+        state.set("Local.obj", _PlainObj())
+
+        result = state.get("Local.obj.__class__.__init__.__globals__.os.environ")
+
+        assert result is None
+        assert sentinel not in str(result)
+
+    def test_rejects_leading_underscore_via_attribute_access(self, state: DeclarativeWorkflowState) -> None:
+        state.set("Local.obj", _PlainObj())
+        assert state.get("Local.obj._private") is None
+
+    def test_rejects_invalid_chars_via_attribute_access(self, state: DeclarativeWorkflowState) -> None:
+        state.set("Local.obj", _PlainObj())
+        assert state.get("Local.obj.text bar") is None
+        assert state.get("Local.obj.text-bar") is None
+
+    def test_rejects_empty_path_and_empty_segments(self, state: DeclarativeWorkflowState) -> None:
+        assert state.get("") is None
+        assert state.get(".") is None
+        assert state.get("Local.") is None
+        assert state.get(".Local") is None
+
+    def test_warning_logged_on_rejected_attribute_segment(
+        self,
+        state: DeclarativeWorkflowState,
+        caplog: pytest.LogCaptureFixture,
+    ) -> None:
+        state.set("Local.obj", _PlainObj())
+        with caplog.at_level(logging.WARNING, logger="agent_framework_declarative._workflows._declarative_base"):
+            state.get("Local.obj.__class__")
+        assert any("rejecting attribute segment" in r.message for r in caplog.records)
+
+    def test_dict_keyed_dunder_is_not_attribute_access(self, state: DeclarativeWorkflowState) -> None:
+        """A literal dunder dict key is harmless because dict lookup never reaches getattr."""
+        state.set("Local.bag", {"__class__": "harmless-string"})
+        assert state.get("Local.bag.__class__") == "harmless-string"
+
+
+# ---------------------------------------------------------------------------
+# get(): legitimate paths continue to work
+# ---------------------------------------------------------------------------
+
+
+class TestGetAllowsValidPaths:
+    def test_underscore_inside_identifier(self, state: DeclarativeWorkflowState) -> None:
+        state.set("Local.user_input", "ok")
+        assert state.get("Local.user_input") == "ok"
+
+    def test_mixed_case_identifiers(self, state: DeclarativeWorkflowState) -> None:
+        state.set("Local.UserInput", "u1")
+        state.set("Local.userInput", "u2")
+        assert state.get("Local.UserInput") == "u1"
+        assert state.get("Local.userInput") == "u2"
+
+    def test_object_attribute_traversal_still_works(self, state: DeclarativeWorkflowState) -> None:
+        state.set("Local.msg", _PlainObj(text="hello"))
+        assert state.get("Local.msg.text") == "hello"
+
+    def test_nested_dict_traversal_still_works(self, state: DeclarativeWorkflowState) -> None:
+        state.set("Local.params", {"team": {"name": "alpha"}})
+        assert state.get("Local.params.team.name") == "alpha"
+
+    def test_uuid_and_hyphenated_dict_keys_are_allowed(self, state: DeclarativeWorkflowState) -> None:
+        """Conversation-id style paths use arbitrary dict keys (UUIDs / hyphens)."""
+        conv_id = "eb815014-06f1-4db6-b7c1-304ea135424f"
+        state.set(f"System.conversations.{conv_id}.messages", ["m1", "m2"])
+        assert state.get(f"System.conversations.{conv_id}.messages") == ["m1", "m2"]
+
+
+# ---------------------------------------------------------------------------
+# set() / append(): dict-keyed operations accept arbitrary string keys
+# ---------------------------------------------------------------------------
+
+
+class TestSetAndAppend:
+    def test_set_allows_underscore_inside_identifier(self, state: DeclarativeWorkflowState) -> None:
+        state.set("Local.user_input", "ok")
+        assert state.get("Local.user_input") == "ok"
+
+    def test_set_allows_uuid_and_hyphenated_dict_keys(self, state: DeclarativeWorkflowState) -> None:
+        conv_id = "conv-test-1"
+        state.set(f"System.conversations.{conv_id}.messages", [])
+        assert state.get(f"System.conversations.{conv_id}.messages") == []
+
+    def test_append_allows_uuid_and_hyphenated_dict_keys(self, state: DeclarativeWorkflowState) -> None:
+        conv_id = "conv-42"
+        state.append(f"System.conversations.{conv_id}.messages", {"role": "user", "text": "hi"})
+        msgs = state.get(f"System.conversations.{conv_id}.messages")
+        assert msgs == [{"role": "user", "text": "hi"}]
+
+    def test_workflow_inputs_still_read_only(self, state: DeclarativeWorkflowState) -> None:
+        with pytest.raises(ValueError, match="read-only"):
+            state.set("Workflow.Inputs.x", 1)
+
+
+# ---------------------------------------------------------------------------
+# set() / append(): malformed paths (empty segments) raise ValueError
+# ---------------------------------------------------------------------------
+
+
+class TestSetRejectsInvalidPaths:
+    @pytest.mark.parametrize("bad_path", ["", "Local.", "Local..foo", ".Local"])
+    def test_set_rejects_empty_segment(self, state: DeclarativeWorkflowState, bad_path: str) -> None:
+        with pytest.raises(ValueError, match="empty segments are not allowed"):
+            state.set(bad_path, "x")
+
+    @pytest.mark.parametrize("bad_path", ["", "Local.", "Local..foo", ".Local"])
+    def test_append_rejects_empty_segment(self, state: DeclarativeWorkflowState, bad_path: str) -> None:
+        with pytest.raises(ValueError, match="empty segments are not allowed"):
+            state.append(bad_path, "x")
+
+    def test_set_rejection_makes_no_partial_write(self, state: DeclarativeWorkflowState) -> None:
+        """Rejected set() must not create an unreachable entry in the state."""
+        state.set("Local.user_input", "pre")
+        with pytest.raises(ValueError):
+            state.set("Local.", "value")
+        local = state.get_state_data().get("Local", {})
+        assert "" not in local
+        assert local == {"user_input": "pre"}
+        assert state.get("Local.") is None
+        assert state.get("Local.user_input") == "pre"
+
+    def test_append_rejection_makes_no_partial_write(self, state: DeclarativeWorkflowState) -> None:
+        """Rejected append() must not create an unreachable entry in the state."""
+        state.set("Local.items", ["a"])
+        with pytest.raises(ValueError):
+            state.append("Local.", "value")
+        local = state.get_state_data().get("Local", {})
+        assert "" not in local
+        assert local == {"items": ["a"]}
+
+
+# ---------------------------------------------------------------------------
+# interpolate_string(): permissive matcher; get() enforces safety
+# ---------------------------------------------------------------------------
+
+
+class TestInterpolateString:
+    def test_ignores_dunder_payload(self, state: DeclarativeWorkflowState, monkeypatch) -> None:
+        sentinel = "agent-framework-interp-sentinel"
+        monkeypatch.setenv("AF_INTERP_SENTINEL", sentinel)
+        state.set("Local.obj", _PlainObj())
+
+        out = state.interpolate_string("X={Local.obj.__class__.__init__.__globals__.os.environ}")
+
+        assert sentinel not in out
+        assert out == "X="
+
+    def test_unknown_path_reduces_to_empty(self, state: DeclarativeWorkflowState) -> None:
+        assert state.interpolate_string("v={Local._private}") == "v="
+
+    @pytest.mark.parametrize(
+        "literal",
+        ["{foo-bar}", "{Ctrl+C}", "{not:a:path}", "{Local.}", "{}"],
+    )
+    def test_non_state_braced_tokens_left_literal(self, state: DeclarativeWorkflowState, literal: str) -> None:
+        assert state.interpolate_string(f"v={literal}") == f"v={literal}"
+
+    def test_allows_underscore_inside_identifier(self, state: DeclarativeWorkflowState) -> None:
+        state.set("Local.user_input", "hello")
+        assert state.interpolate_string("v={Local.user_input}") == "v=hello"
+
+    def test_resolves_nested_dict_path(self, state: DeclarativeWorkflowState) -> None:
+        state.set("Local.params", {"team": "alpha"})
+        assert state.interpolate_string("team={Local.params.team}") == "team=alpha"
+
+    @pytest.mark.parametrize(
+        ("key", "value"),
+        [
+            ("_id", "abc123"),
+            ("1", "one"),
+            ("2025", "year-bucket"),
+        ],
+    )
+    def test_resolves_dict_keyed_segments(self, state: DeclarativeWorkflowState, key: str, value: str) -> None:
+        state.set("Local.bag", {key: value})
+        assert state.interpolate_string(f"v={{Local.bag.{key}}}") == f"v={value}"
+
+    def test_resolves_uuid_conversation_key(self, state: DeclarativeWorkflowState) -> None:
+        conv_id = "eb815014-06f1-4db6-b7c1-304ea135424f"
+        state.set(f"System.conversations.{conv_id}.messages", ["hello"])
+        out = state.interpolate_string(f"m={{System.conversations.{conv_id}.messages}}")
+        assert out == "m=['hello']"
+
+    def test_end_to_end_send_activity_payload_neutralized(
+        self,
+        state: DeclarativeWorkflowState,
+        monkeypatch,
+    ) -> None:
+        sentinel = "agent-framework-e2e-sentinel"
+        monkeypatch.setenv("AF_E2E_SENTINEL", sentinel)
+        state.set("Local.toolResult", _PlainObj())
+
+        payload = "{Local.toolResult.__class__.__init__.__globals__.os.environ}"
+        evaluated = state.eval_if_expression(payload)
+        rendered = state.interpolate_string(evaluated) if isinstance(evaluated, str) else str(evaluated)
+
+        assert sentinel not in rendered
+        assert rendered == ""
+
+
+# ---------------------------------------------------------------------------
+# Regressions: PowerFx and internal temp-variable handling still work
+# ---------------------------------------------------------------------------
+
+
+@_requires_powerfx
+class TestPowerFxStillWorks:
+    def test_simple_powerfx_expression_evaluates(self, state: DeclarativeWorkflowState) -> None:
+        state.set("Local.x", 6)
+        state.set("Local.y", 7)
+        assert state.eval("=Local.x * Local.y") == 42
+
+    def test_internal_temp_message_text_still_works(self, state: DeclarativeWorkflowState) -> None:
+        """Long MessageText() results round-trip and the temp key is removed after eval."""
+        long_text = "A" * 600
+        state.set(
+            "Local.Messages",
+            [{"text": long_text, "contents": [{"type": "text", "text": long_text}]}],
+        )
+
+        result = state.eval("=Upper(MessageText(Local.Messages))")
+        assert result == "A" * 600
+
+        local = state.get_state_data().get("Local", {})
+        remaining = sorted(k for k in local if k.startswith("_TempMessageText"))
+        assert not remaining, f"Temporary keys remain in Local: {remaining}"
+
+    def test_message_text_eval_preserves_user_temp_value(self, state: DeclarativeWorkflowState) -> None:
+        """User state at the temp key path survives a long MessageText eval."""
+        long_text = "A" * 600
+        state.set("Local._TempMessageText0", "user-important-value")
+        state.set(
+            "Local.Messages",
+            [{"text": long_text, "contents": [{"type": "text", "text": long_text}]}],
+        )
+
+        result = state.eval("=Upper(MessageText(Local.Messages))")
+        assert result == "A" * 600
+        assert state.get("Local._TempMessageText0") == "user-important-value"
+
+    def test_message_text_eval_cleans_up_on_powerfx_failure(
+        self,
+        state: DeclarativeWorkflowState,
+        monkeypatch,
+    ) -> None:
+        """Temp key is removed even when PowerFx evaluation raises."""
+        from agent_framework_declarative._workflows import _declarative_base as base
+
+        class _FailingEngine:
+            def eval(self, *args: Any, **kwargs: Any) -> Any:
+                raise RuntimeError("boom")
+
+        monkeypatch.setattr(base, "Engine", _FailingEngine)
+
+        long_text = "A" * 600
+        state.set(
+            "Local.Messages",
+            [{"text": long_text, "contents": [{"type": "text", "text": long_text}]}],
+        )
+
+        with pytest.raises(RuntimeError, match="boom"):
+            state.eval("=Upper(MessageText(Local.Messages))")
+
+        local = state.get_state_data().get("Local", {})
+        remaining = sorted(k for k in local if k.startswith("_TempMessageText"))
+        assert not remaining, f"Temporary keys remain in Local after PowerFx failure: {remaining}"

python/packages/declarative/tests/test_function_tool_executor.py

@@ -35,14 +35,12 @@ pytestmark = pytest.mark.skipif(
 from agent_framework_declarative._workflows import (  # noqa: E402
     DECLARATIVE_STATE_KEY,
     FUNCTION_TOOL_REGISTRY_KEY,
-    TOOL_APPROVAL_STATE_KEY,
     ActionComplete,
     ActionTrigger,
     DeclarativeWorkflowBuilder,
     InvokeFunctionToolExecutor,
     ToolApprovalRequest,
     ToolApprovalResponse,
-    ToolApprovalState,
     ToolInvocationResult,
     WorkflowFactory,
 )
@@ -393,21 +391,6 @@ class TestToolApprovalTypes:
         assert response.approved is False
         assert response.reason == "Not authorized"
 
-    def test_approval_state(self):
-        """Test creating approval state for yield/resume."""
-        state = ToolApprovalState(
-            function_name="delete_user",
-            arguments={"user_id": "123"},
-            output_messages_var="Local.messages",
-            output_result_var="Local.result",
-            auto_send=True,
-        )
-        assert state.function_name == "delete_user"
-        assert state.arguments == {"user_id": "123"}
-        assert state.output_messages_var == "Local.messages"
-        assert state.output_result_var == "Local.result"
-        assert state.auto_send is True
-
 
 class TestInvokeFunctionToolEdgeCases:
     """Tests for edge cases and error handling."""
@@ -1075,13 +1058,6 @@ class TestApprovalFlow:
         # Should NOT have sent ActionComplete (workflow yields)
         mock_context.send_message.assert_not_called()
 
-        # Approval state should be saved in state
-        approval_key = f"{TOOL_APPROVAL_STATE_KEY}_approval_test"
-        saved_state = mock_state._data[approval_key]
-        assert isinstance(saved_state, ToolApprovalState)
-        assert saved_state.function_name == "my_tool"
-        assert saved_state.arguments == {"x": 5}
-
     @pytest.mark.asyncio
     async def test_approval_response_approved(self, mock_state, mock_context):
         """When approval response is approved, the tool should be invoked."""
@@ -1104,17 +1080,7 @@ class TestApprovalFlow:
 
         executor = InvokeFunctionToolExecutor(action_def, tools={"my_tool": my_tool})
 
-        # Pre-populate approval state (simulating what handle_action stores)
-        approval_key = f"{TOOL_APPROVAL_STATE_KEY}_approval_approved"
-        mock_state._data[approval_key] = ToolApprovalState(
-            function_name="my_tool",
-            arguments={"x": 7},
-            output_messages_var=None,
-            output_result_var="Local.result",
-            auto_send=True,
-        )
-
-        # Simulate the response
+        # Simulate the response — invocation params come from original_request
         original_request = ToolApprovalRequest(
             request_id="req-123",
             function_name="my_tool",
@@ -1124,7 +1090,7 @@ class TestApprovalFlow:
 
         await executor.handle_approval_response(original_request, response, mock_context)
 
-        # Tool should have been called
+        # Tool should have been called with the approved arguments
         assert call_log == [7]
 
         # ActionComplete should have been sent
@@ -1132,9 +1098,6 @@ class TestApprovalFlow:
         sent = mock_context.send_message.call_args[0][0]
         assert isinstance(sent, ActionComplete)
 
-        # Approval state should be cleaned up
-        assert approval_key not in mock_state._data
-
     @pytest.mark.asyncio
     async def test_approval_response_rejected(self, mock_state, mock_context):
         """When approval response is rejected, rejection status should be stored."""
@@ -1154,16 +1117,6 @@ class TestApprovalFlow:
 
         executor = InvokeFunctionToolExecutor(action_def, tools={"my_tool": my_tool})
 
-        # Pre-populate approval state
-        approval_key = f"{TOOL_APPROVAL_STATE_KEY}_approval_rejected"
-        mock_state._data[approval_key] = ToolApprovalState(
-            function_name="my_tool",
-            arguments={"x": 5},
-            output_messages_var=None,
-            output_result_var="Local.result",
-            auto_send=True,
-        )
-
         original_request = ToolApprovalRequest(
             request_id="req-456",
             function_name="my_tool",
@@ -1185,36 +1138,6 @@ class TestApprovalFlow:
         assert result["reason"] == "Not authorized"
         assert result["approved"] is False
 
-    @pytest.mark.asyncio
-    async def test_approval_response_missing_state(self, mock_state, mock_context):
-        """When approval state is missing on resume, should log error and complete."""
-        self._init_state(mock_state)
-
-        action_def = {
-            "kind": "InvokeFunctionTool",
-            "id": "missing_state_test",
-            "functionName": "my_tool",
-            "requireApproval": True,
-            "output": {"result": "Local.result"},
-        }
-
-        executor = InvokeFunctionToolExecutor(action_def, tools={})
-
-        # Don't populate approval state - simulate missing state
-        original_request = ToolApprovalRequest(
-            request_id="req-789",
-            function_name="my_tool",
-            arguments={},
-        )
-        response = ToolApprovalResponse(approved=True)
-
-        await executor.handle_approval_response(original_request, response, mock_context)
-
-        # Should still send ActionComplete
-        mock_context.send_message.assert_called_once()
-        sent = mock_context.send_message.call_args[0][0]
-        assert isinstance(sent, ActionComplete)
-
 
 # ============================================================================
 # State registry tool lookup (lines 255-257)

python/packages/declarative/tests/test_graph_coverage.py

@@ -2765,7 +2765,7 @@ class TestLongMessageTextHandling:
         assert temp_var is None
 
     async def test_long_message_text_stored_in_temp_variable(self, mock_state):
-        """Test that long MessageText results are stored in temp variables."""
+        """Long MessageText results round-trip and the temp key is removed after eval."""
         state = DeclarativeWorkflowState(mock_state)
         state.initialize()
 
@@ -2777,9 +2777,9 @@ class TestLongMessageTextHandling:
         result = state.eval("=Upper(MessageText(Local.Messages))")
         assert result == "A" * 600  # Upper on 'A' is still 'A'
 
-        # A temp variable should have been created
-        temp_var = state.get("Local._TempMessageText0")
-        assert temp_var == long_text
+        local = state.get_state_data().get("Local", {})
+        remaining = sorted(k for k in local if k.startswith("_TempMessageText"))
+        assert not remaining, f"Temporary keys remain in Local: {remaining}"
 
     async def test_find_with_long_message_text(self, mock_state):
         """Test Find function works with long MessageText stored in temp variable."""

python/packages/declarative/tests/test_invoke_mcp_tool_executor.py

@@ -403,7 +403,6 @@ class TestApprovalFlow:
     async def test_approval_required_emits_request_and_yields(self, mock_state, mock_context) -> None:  # type: ignore[no-untyped-def]
         from agent_framework_declarative._workflows._declarative_base import ActionTrigger
         from agent_framework_declarative._workflows._executors_mcp import (
-            _MCP_APPROVAL_STATE_KEY,
             InvokeMcpToolActionExecutor,
             MCPToolApprovalRequest,
         )
@@ -439,18 +438,12 @@ class TestApprovalFlow:
         # Handler not invoked yet.
         assert handler.call_count == 0
 
-        # Approval state stored.
-        approval_key = f"{_MCP_APPROVAL_STATE_KEY}_mcp_action"
-        assert approval_key in mock_state._data
-
     @pytest.mark.asyncio
     async def test_approval_response_approved_invokes_handler(self, mock_state, mock_context) -> None:  # type: ignore[no-untyped-def]
         from agent_framework_declarative._workflows import ActionComplete, ToolApprovalResponse
         from agent_framework_declarative._workflows._executors_mcp import (
-            _MCP_APPROVAL_STATE_KEY,
             InvokeMcpToolActionExecutor,
             MCPToolApprovalRequest,
-            _MCPToolApprovalState,
         )
 
         _seed_state(mock_state)
@@ -458,24 +451,11 @@ class TestApprovalFlow:
         executor = InvokeMcpToolActionExecutor(
             _action(
                 require_approval=True,
+                headers={"Authorization": "Bearer tk"},
                 output={"result": "Local.Result"},
             ),
             mcp_tool_handler=handler,
         )
-        # Pre-populate approval state.
-        approval_key = f"{_MCP_APPROVAL_STATE_KEY}_mcp_action"
-        mock_state._data[approval_key] = _MCPToolApprovalState(
-            server_url="https://mcp.example/api",
-            tool_name="search",
-            server_label=None,
-            arguments={"q": "x"},
-            connection_name=None,
-            headers_def={"Authorization": "Bearer tk"},
-            auto_send=False,
-            conversation_id_expr=None,
-            output_messages_path=None,
-            output_result_path="Local.Result",
-        )
         await executor.handle_approval_response(
             MCPToolApprovalRequest(
                 request_id="req-1",
@@ -491,10 +471,12 @@ class TestApprovalFlow:
         assert handler.call_count == 1
         inv = handler.last_invocation
         assert inv is not None
-        # Headers are re-evaluated from headers_def.
+        # Invocation fields source from the approval request payload.
+        assert inv.tool_name == "search"
+        assert inv.server_url == "https://mcp.example/api"
+        assert inv.arguments == {"q": "x"}
+        # Headers are re-evaluated from the action definition on resume.
         assert inv.headers == {"Authorization": "Bearer tk"}
-        # Approval state was cleaned up.
-        assert approval_key not in mock_state._data
         # ActionComplete was sent.
         mock_context.send_message.assert_called_once()
         sent = mock_context.send_message.call_args[0][0]
@@ -504,10 +486,8 @@ class TestApprovalFlow:
     async def test_approval_response_rejected_assigns_error(self, mock_state, mock_context) -> None:  # type: ignore[no-untyped-def]
         from agent_framework_declarative._workflows import ToolApprovalResponse
         from agent_framework_declarative._workflows._executors_mcp import (
-            _MCP_APPROVAL_STATE_KEY,
             InvokeMcpToolActionExecutor,
             MCPToolApprovalRequest,
-            _MCPToolApprovalState,
         )
 
         _seed_state(mock_state)
@@ -519,19 +499,6 @@ class TestApprovalFlow:
             ),
             mcp_tool_handler=handler,
         )
-        approval_key = f"{_MCP_APPROVAL_STATE_KEY}_mcp_action"
-        mock_state._data[approval_key] = _MCPToolApprovalState(
-            server_url="https://mcp.example/api",
-            tool_name="search",
-            server_label=None,
-            arguments={},
-            connection_name=None,
-            headers_def=None,
-            auto_send=True,
-            conversation_id_expr=None,
-            output_messages_path=None,
-            output_result_path="Local.Result",
-        )
         await executor.handle_approval_response(
             MCPToolApprovalRequest(
                 request_id="req-2",

python/packages/openai/agent_framework_openai/_chat_client.py

@@ -2979,10 +2979,16 @@ class RawOpenAIChatClient(  # type: ignore[misc]
             output_token_count=usage.output_tokens,
             total_token_count=usage.total_tokens,
         )
-        if usage.input_tokens_details and usage.input_tokens_details.cached_tokens:
-            details["openai.cached_input_tokens"] = usage.input_tokens_details.cached_tokens  # type: ignore[typeddict-unknown-key]
-        if usage.output_tokens_details and usage.output_tokens_details.reasoning_tokens:
-            details["openai.reasoning_tokens"] = usage.output_tokens_details.reasoning_tokens  # type: ignore[typeddict-unknown-key]
+        if usage.input_tokens_details:
+            cached_tokens = cast("int | None", getattr(usage.input_tokens_details, "cached_tokens", None))
+            if cached_tokens is not None:
+                details["openai.cached_input_tokens"] = cached_tokens  # type: ignore[typeddict-unknown-key]
+                details["cache_read_input_token_count"] = cached_tokens
+        if usage.output_tokens_details:
+            reasoning_tokens = cast("int | None", getattr(usage.output_tokens_details, "reasoning_tokens", None))
+            if reasoning_tokens is not None:
+                details["openai.reasoning_tokens"] = reasoning_tokens  # type: ignore[typeddict-unknown-key]
+                details["reasoning_output_token_count"] = reasoning_tokens
         return details
 
     def _get_metadata_from_response(self, output: Any) -> dict[str, Any]:

python/packages/openai/agent_framework_openai/_chat_completion_client.py

@@ -765,15 +765,17 @@ class RawOpenAIChatCompletionClient(  # type: ignore[misc]
                 details["completion/accepted_prediction_tokens"] = tokens  # type: ignore[typeddict-unknown-key]
             if tokens := usage.completion_tokens_details.audio_tokens:
                 details["completion/audio_tokens"] = tokens  # type: ignore[typeddict-unknown-key]
-            if tokens := usage.completion_tokens_details.reasoning_tokens:
+            if (tokens := usage.completion_tokens_details.reasoning_tokens) is not None:
                 details["completion/reasoning_tokens"] = tokens  # type: ignore[typeddict-unknown-key]
+                details["reasoning_output_token_count"] = tokens
             if tokens := usage.completion_tokens_details.rejected_prediction_tokens:
                 details["completion/rejected_prediction_tokens"] = tokens  # type: ignore[typeddict-unknown-key]
         if usage.prompt_tokens_details:
             if tokens := usage.prompt_tokens_details.audio_tokens:
                 details["prompt/audio_tokens"] = tokens  # type: ignore[typeddict-unknown-key]
-            if tokens := usage.prompt_tokens_details.cached_tokens:
+            if (tokens := usage.prompt_tokens_details.cached_tokens) is not None:
                 details["prompt/cached_tokens"] = tokens  # type: ignore[typeddict-unknown-key]
+                details["cache_read_input_token_count"] = tokens
         return details
 
     def _parse_text_from_openai(self, choice: Choice | ChunkChoice) -> Content | None:

python/packages/openai/tests/openai/test_openai_chat_client.py

@@ -3301,6 +3301,7 @@ def test_usage_details_with_cached_tokens() -> None:
     assert details is not None
     assert details["input_token_count"] == 200
     assert details["openai.cached_input_tokens"] == 25
+    assert details["cache_read_input_token_count"] == 25
 
 
 def test_usage_details_with_reasoning_tokens() -> None:
@@ -3319,6 +3320,49 @@ def test_usage_details_with_reasoning_tokens() -> None:
     assert details is not None
     assert details["output_token_count"] == 80
     assert details["openai.reasoning_tokens"] == 30
+    assert details["reasoning_output_token_count"] == 30
+
+
+def test_usage_details_with_zero_cached_and_reasoning_tokens() -> None:
+    """Test _parse_usage_from_openai preserves zero-valued mapped usage details."""
+    client = OpenAIChatClient(model="test-model", api_key="test-key")
+
+    mock_usage = MagicMock()
+    mock_usage.input_tokens = 150
+    mock_usage.output_tokens = 80
+    mock_usage.total_tokens = 230
+    mock_usage.input_tokens_details = MagicMock()
+    mock_usage.input_tokens_details.cached_tokens = 0
+    mock_usage.output_tokens_details = MagicMock()
+    mock_usage.output_tokens_details.reasoning_tokens = 0
+
+    details = client._parse_usage_from_openai(mock_usage)  # type: ignore
+    assert details is not None
+    assert details["openai.cached_input_tokens"] == 0
+    assert details["cache_read_input_token_count"] == 0
+    assert details["openai.reasoning_tokens"] == 0
+    assert details["reasoning_output_token_count"] == 0
+
+
+def test_usage_details_omits_missing_cached_and_reasoning_tokens() -> None:
+    """Test _parse_usage_from_openai omits missing mapped usage details."""
+    client = OpenAIChatClient(model="test-model", api_key="test-key")
+
+    mock_usage = MagicMock()
+    mock_usage.input_tokens = 150
+    mock_usage.output_tokens = 80
+    mock_usage.total_tokens = 230
+    mock_usage.input_tokens_details = MagicMock()
+    mock_usage.input_tokens_details.cached_tokens = None
+    mock_usage.output_tokens_details = MagicMock()
+    mock_usage.output_tokens_details.reasoning_tokens = None
+
+    details = client._parse_usage_from_openai(mock_usage)  # type: ignore
+    assert details is not None
+    assert "openai.cached_input_tokens" not in details
+    assert "cache_read_input_token_count" not in details
+    assert "openai.reasoning_tokens" not in details
+    assert "reasoning_output_token_count" not in details
 
 
 def test_get_metadata_from_response() -> None:

python/packages/openai/tests/openai/test_openai_chat_completion_client.py

@@ -1099,6 +1099,31 @@ def test_usage_content_in_streaming_response(
     assert usage_content.usage_details["total_token_count"] == 150
 
 
+def test_parse_usage_includes_standard_and_legacy_mapped_token_details() -> None:
+    """Test _parse_usage_from_openai emits standard and legacy mapped token details."""
+    client = OpenAIChatCompletionClient(model="test-model", api_key="test-key")
+
+    mock_usage = MagicMock()
+    mock_usage.prompt_tokens = 100
+    mock_usage.completion_tokens = 50
+    mock_usage.total_tokens = 150
+    mock_usage.completion_tokens_details = MagicMock()
+    mock_usage.completion_tokens_details.accepted_prediction_tokens = None
+    mock_usage.completion_tokens_details.audio_tokens = None
+    mock_usage.completion_tokens_details.reasoning_tokens = 0
+    mock_usage.completion_tokens_details.rejected_prediction_tokens = None
+    mock_usage.prompt_tokens_details = MagicMock()
+    mock_usage.prompt_tokens_details.audio_tokens = None
+    mock_usage.prompt_tokens_details.cached_tokens = 0
+
+    details = client._parse_usage_from_openai(mock_usage)  # type: ignore[arg-type]
+
+    assert details["completion/reasoning_tokens"] == 0
+    assert details["reasoning_output_token_count"] == 0
+    assert details["prompt/cached_tokens"] == 0
+    assert details["cache_read_input_token_count"] == 0
+
+
 def test_streaming_chunk_with_usage_and_text(
     openai_unit_test_env: dict[str, str],
 ) -> None: