Python: DevUI Fix Serialization, Timestamp and Other Issues (#1584)

* refactor(devui): adopt standard OpenAI lifecycle events for agents and workflows

- Replace custom workflow events with OpenAI Responses API standard lifecycle events
- Add AgentStartedEvent, AgentCompletedEvent, AgentFailedEvent for clean separation
- Implement ExecutorActionItem for workflow executor tracking
- Convert informational events to trace events to reduce noise
- Update README mapper table with comprehensive event mappings
- Maintain full backward compatibility with legacy events

* fix(devui): resolve timestamp overwriting and Content serialization errors

- Fix tool call timestamps being overwritten on each render (#1483)
- Add recursive Content serialization to handle ChatMessage and nested objects (#1548)
- Implement proper MCP tool cleanup on server shutdown
- Add timestamp field to function_result.complete events
- Enhance credential and client resource cleanup

Fixes #1483, #1548
Partial improvements for #1476

python/packages/devui/agent_framework_devui/_discovery.py

@@ -127,7 +127,7 @@ class EntityDiscovery:
 
         # Cache the loaded object
         self._loaded_objects[entity_id] = entity_obj
-        logger.info(f"✅ Successfully loaded entity: {entity_id} (type: {enriched_info.type})")
+        logger.info(f"Successfully loaded entity: {entity_id} (type: {enriched_info.type})")
 
         return entity_obj
 
@@ -217,7 +217,7 @@ class EntityDiscovery:
         if entity_info and "lazy_loaded" in entity_info.metadata:
             entity_info.metadata["lazy_loaded"] = False
 
-        logger.info(f"♻️  Entity invalidated: {entity_id} (will reload on next access)")
+        logger.info(f"Entity invalidated: {entity_id} (will reload on next access)")
 
     def invalidate_all(self) -> None:
         """Invalidate all cached entities.

python/packages/devui/agent_framework_devui/_executor.py

@@ -217,6 +217,11 @@ class AgentFrameworkExecutor:
             Agent update events and trace events
         """
         try:
+            # Emit agent lifecycle start event
+            from .models._openai_custom import AgentStartedEvent
+
+            yield AgentStartedEvent()
+
             # Convert input to proper ChatMessage or string
             user_message = self._convert_input_to_chat_message(request.input)
 
@@ -266,8 +271,19 @@ class AgentFrameworkExecutor:
             else:
                 raise ValueError("Agent must implement either run() or run_stream() method")
 
+            # Emit agent lifecycle completion event
+            from .models._openai_custom import AgentCompletedEvent
+
+            yield AgentCompletedEvent()
+
         except Exception as e:
             logger.error(f"Error in agent execution: {e}")
+            # Emit agent lifecycle failure event
+            from .models._openai_custom import AgentFailedEvent
+
+            yield AgentFailedEvent(error=e)
+
+            # Still yield the error for backward compatibility
             yield {"type": "error", "message": f"Agent execution error: {e!s}"}
 
     async def _execute_workflow(
@@ -284,14 +300,9 @@ class AgentFrameworkExecutor:
             Workflow events and trace events
         """
         try:
-            # Get input data - prefer structured data from extra_body
-            input_data: str | list[Any] | dict[str, Any]
-            if request.extra_body and isinstance(request.extra_body, dict) and request.extra_body.get("input_data"):
-                input_data = request.extra_body.get("input_data")  # type: ignore
-                logger.debug(f"Using structured input_data from extra_body: {type(input_data)}")
-            else:
-                input_data = request.input
-                logger.debug(f"Using input field as fallback: {type(input_data)}")
+            # Get input data directly from request.input field
+            input_data = request.input
+            logger.debug(f"Using input field: {type(input_data)}")
 
             # Parse input based on workflow's expected input type
             parsed_input = await self._parse_workflow_input(workflow, input_data)

python/packages/devui/agent_framework_devui/_mapper.py

@@ -4,17 +4,32 @@
 
 import json
 import logging
+import time
 import uuid
 from collections import OrderedDict
 from collections.abc import Sequence
 from datetime import datetime
 from typing import Any, Union
+from uuid import uuid4
+
+from openai.types.responses import (
+    Response,
+    ResponseContentPartAddedEvent,
+    ResponseCreatedEvent,
+    ResponseError,
+    ResponseFailedEvent,
+    ResponseInProgressEvent,
+)
 
 from .models import (
     AgentFrameworkRequest,
+    CustomResponseOutputItemAddedEvent,
+    CustomResponseOutputItemDoneEvent,
+    ExecutorActionItem,
     InputTokensDetails,
     OpenAIResponse,
     OutputTokensDetails,
+    ResponseCompletedEvent,
     ResponseErrorEvent,
     ResponseFunctionCallArgumentsDeltaEvent,
     ResponseFunctionResultComplete,
@@ -41,6 +56,56 @@ EventType = Union[
 ]
 
 
+def _serialize_content_recursive(value: Any) -> Any:
+    """Recursively serialize Agent Framework Content objects to JSON-compatible values.
+
+    This handles nested Content objects (like TextContent inside FunctionResultContent.result)
+    that can't be directly serialized by json.dumps().
+
+    Args:
+        value: Value to serialize (can be Content object, dict, list, primitive, etc.)
+
+    Returns:
+        JSON-serializable version with all Content objects converted to dicts/primitives
+    """
+    # Handle None and basic JSON-serializable types
+    if value is None or isinstance(value, (str, int, float, bool)):
+        return value
+
+    # Check if it's a SerializationMixin (includes all Content types)
+    # Content objects have to_dict() method
+    if hasattr(value, "to_dict") and callable(getattr(value, "to_dict", None)):
+        try:
+            return value.to_dict()
+        except Exception as e:
+            # If to_dict() fails, fall through to other methods
+            logger.debug(f"Failed to serialize with to_dict(): {e}")
+
+    # Handle dictionaries - recursively process values
+    if isinstance(value, dict):
+        return {key: _serialize_content_recursive(val) for key, val in value.items()}
+
+    # Handle lists and tuples - recursively process elements
+    if isinstance(value, (list, tuple)):
+        serialized = [_serialize_content_recursive(item) for item in value]
+        # For single-item lists containing text Content, extract just the text
+        # This handles the MCP case where result = [TextContent(text="Hello")]
+        # and we want output = "Hello" not output = '[{"type": "text", "text": "Hello"}]'
+        if len(serialized) == 1 and isinstance(serialized[0], dict) and serialized[0].get("type") == "text":
+            return serialized[0].get("text", "")
+        return serialized
+
+    # For other objects with model_dump(), try that
+    if hasattr(value, "model_dump") and callable(getattr(value, "model_dump", None)):
+        try:
+            return value.model_dump()
+        except Exception as e:
+            logger.debug(f"Failed to serialize with model_dump(): {e}")
+
+    # Return as-is and let json.dumps handle it (may raise TypeError for non-serializable types)
+    return value
+
+
 class MessageMapper:
     """Maps Agent Framework messages/responses to OpenAI format."""
 
@@ -102,6 +167,12 @@ class MessageMapper:
                 )
             ]
 
+        # Handle Agent lifecycle events first
+        from .models._openai_custom import AgentCompletedEvent, AgentFailedEvent, AgentStartedEvent
+
+        if isinstance(raw_event, (AgentStartedEvent, AgentCompletedEvent, AgentFailedEvent)):
+            return await self._convert_agent_lifecycle_event(raw_event, context)
+
         # Import Agent Framework types for proper isinstance checks
         try:
             from agent_framework import AgentRunResponse, AgentRunResponseUpdate, WorkflowEvent
@@ -245,6 +316,7 @@ class MessageMapper:
                 "content_index": 0,
                 "output_index": 0,
                 "request_id": str(request_key),  # For usage accumulation
+                "request": request,  # Store the request for model name access
                 # Track active function calls: {call_id: {name, item_id, args_chunks}}
                 "active_function_calls": {},
             }
@@ -267,7 +339,7 @@ class MessageMapper:
         return int(context["sequence_counter"])
 
     async def _convert_agent_update(self, update: Any, context: dict[str, Any]) -> Sequence[Any]:
-        """Convert AgentRunResponseUpdate to OpenAI events using comprehensive content mapping.
+        """Convert agent text updates to proper content part events.
 
         Args:
             update: Agent run response update
@@ -283,10 +355,60 @@ class MessageMapper:
             if not hasattr(update, "contents") or not update.contents:
                 return events
 
+            # Check if we're streaming text content
+            has_text_content = any(content.__class__.__name__ == "TextContent" for content in update.contents)
+
+            # If we have text content and haven't created a message yet, create one
+            if has_text_content and "current_message_id" not in context:
+                message_id = f"msg_{uuid4().hex[:8]}"
+                context["current_message_id"] = message_id
+                context["output_index"] = context.get("output_index", -1) + 1
+
+                # Add message output item
+                events.append(
+                    ResponseOutputItemAddedEvent(
+                        type="response.output_item.added",
+                        output_index=context["output_index"],
+                        sequence_number=self._next_sequence(context),
+                        item=ResponseOutputMessage(
+                            type="message", id=message_id, role="assistant", content=[], status="in_progress"
+                        ),
+                    )
+                )
+
+                # Add content part for text
+                context["content_index"] = 0
+                events.append(
+                    ResponseContentPartAddedEvent(
+                        type="response.content_part.added",
+                        output_index=context["output_index"],
+                        content_index=context["content_index"],
+                        item_id=message_id,
+                        sequence_number=self._next_sequence(context),
+                        part=ResponseOutputText(type="output_text", text="", annotations=[]),
+                    )
+                )
+
+            # Process each content item
             for content in update.contents:
                 content_type = content.__class__.__name__
 
-                if content_type in self.content_mappers:
+                # Special handling for TextContent to use proper delta events
+                if content_type == "TextContent" and "current_message_id" in context:
+                    # Stream text content via proper delta events
+                    events.append(
+                        ResponseTextDeltaEvent(
+                            type="response.output_text.delta",
+                            output_index=context["output_index"],
+                            content_index=context.get("content_index", 0),
+                            item_id=context["current_message_id"],
+                            delta=content.text,
+                            logprobs=[],  # We don't have logprobs from Agent Framework
+                            sequence_number=self._next_sequence(context),
+                        )
+                    )
+                elif content_type in self.content_mappers:
+                    # Use existing mappers for other content types
                     mapped_events = await self.content_mappers[content_type](content, context)
                     if mapped_events is not None:  # Handle None returns (e.g., UsageContent)
                         if isinstance(mapped_events, list):
@@ -297,7 +419,9 @@ class MessageMapper:
                     # Graceful fallback for unknown content types
                     events.append(await self._create_unknown_content_event(content, context))
 
-                context["content_index"] += 1
+                # Don't increment content_index for text deltas within the same part
+                if content_type != "TextContent":
+                    context["content_index"] = context.get("content_index", 0) + 1
 
         except Exception as e:
             logger.warning(f"Error converting agent update: {e}")
@@ -358,8 +482,105 @@ class MessageMapper:
 
         return events
 
+    async def _convert_agent_lifecycle_event(self, event: Any, context: dict[str, Any]) -> Sequence[Any]:
+        """Convert agent lifecycle events to OpenAI response events.
+
+        Args:
+            event: AgentStartedEvent, AgentCompletedEvent, or AgentFailedEvent
+            context: Conversion context
+
+        Returns:
+            List of OpenAI response stream events
+        """
+        from .models._openai_custom import AgentCompletedEvent, AgentFailedEvent, AgentStartedEvent
+
+        try:
+            # Get model name from context (the agent name)
+            model_name = context.get("request", {}).model if context.get("request") else "agent"
+
+            if isinstance(event, AgentStartedEvent):
+                execution_id = f"agent_{uuid4().hex[:12]}"
+                context["execution_id"] = execution_id
+
+                # Create Response object
+                response_obj = Response(
+                    id=f"resp_{execution_id}",
+                    object="response",
+                    created_at=float(time.time()),
+                    model=model_name,
+                    output=[],
+                    status="in_progress",
+                    parallel_tool_calls=False,
+                    tool_choice="none",
+                    tools=[],
+                )
+
+                # Emit both created and in_progress events
+                return [
+                    ResponseCreatedEvent(
+                        type="response.created", sequence_number=self._next_sequence(context), response=response_obj
+                    ),
+                    ResponseInProgressEvent(
+                        type="response.in_progress", sequence_number=self._next_sequence(context), response=response_obj
+                    ),
+                ]
+
+            if isinstance(event, AgentCompletedEvent):
+                execution_id = context.get("execution_id", f"agent_{uuid4().hex[:12]}")
+
+                response_obj = Response(
+                    id=f"resp_{execution_id}",
+                    object="response",
+                    created_at=float(time.time()),
+                    model=model_name,
+                    output=[],
+                    status="completed",
+                    parallel_tool_calls=False,
+                    tool_choice="none",
+                    tools=[],
+                )
+
+                return [
+                    ResponseCompletedEvent(
+                        type="response.completed", sequence_number=self._next_sequence(context), response=response_obj
+                    )
+                ]
+
+            if isinstance(event, AgentFailedEvent):
+                execution_id = context.get("execution_id", f"agent_{uuid4().hex[:12]}")
+
+                # Create error object
+                response_error = ResponseError(
+                    message=str(event.error) if event.error else "Unknown error", code="server_error"
+                )
+
+                response_obj = Response(
+                    id=f"resp_{execution_id}",
+                    object="response",
+                    created_at=float(time.time()),
+                    model=model_name,
+                    output=[],
+                    status="failed",
+                    error=response_error,
+                    parallel_tool_calls=False,
+                    tool_choice="none",
+                    tools=[],
+                )
+
+                return [
+                    ResponseFailedEvent(
+                        type="response.failed", sequence_number=self._next_sequence(context), response=response_obj
+                    )
+                ]
+
+            return []
+
+        except Exception as e:
+            logger.warning(f"Error converting agent lifecycle event: {e}")
+            return [await self._create_error_event(str(e), context)]
+
     async def _convert_workflow_event(self, event: Any, context: dict[str, Any]) -> Sequence[Any]:
-        """Convert workflow event to structured OpenAI events.
+        """Convert workflow events to standard OpenAI event objects.
 
         Args:
             event: Workflow event
@@ -369,22 +590,247 @@ class MessageMapper:
             List of OpenAI response stream events
         """
         try:
+            event_class = event.__class__.__name__
+
+            # Response-level events - construct proper OpenAI objects
+            if event_class == "WorkflowStartedEvent":
+                workflow_id = getattr(event, "workflow_id", str(uuid4()))
+                context["workflow_id"] = workflow_id
+
+                # Import Response type for proper construction
+                from openai.types.responses import Response
+
+                # Return proper OpenAI event objects
+                events: list[Any] = []
+
+                # Determine the model name - use request model or default to "workflow"
+                # The request model will be the agent name for agents, workflow name for workflows
+                model_name = context.get("request", {}).model if context.get("request") else "workflow"
+
+                # Create a full Response object with all required fields
+                response_obj = Response(
+                    id=f"resp_{workflow_id}",
+                    object="response",
+                    created_at=float(time.time()),
+                    model=model_name,  # Use the actual model/agent name
+                    output=[],  # Empty output list initially
+                    status="in_progress",
+                    # Required fields with safe defaults
+                    parallel_tool_calls=False,
+                    tool_choice="none",
+                    tools=[],
+                )
+
+                # First emit response.created
+                events.append(
+                    ResponseCreatedEvent(
+                        type="response.created", sequence_number=self._next_sequence(context), response=response_obj
+                    )
+                )
+
+                # Then emit response.in_progress (reuse same response object)
+                events.append(
+                    ResponseInProgressEvent(
+                        type="response.in_progress", sequence_number=self._next_sequence(context), response=response_obj
+                    )
+                )
+
+                return events
+
+            if event_class in ["WorkflowCompletedEvent", "WorkflowOutputEvent"]:
+                workflow_id = context.get("workflow_id", str(uuid4()))
+
+                # Import Response type for proper construction
+                from openai.types.responses import Response
+
+                # Get model name from context
+                model_name = context.get("request", {}).model if context.get("request") else "workflow"
+
+                # Create a full Response object for completed state
+                response_obj = Response(
+                    id=f"resp_{workflow_id}",
+                    object="response",
+                    created_at=float(time.time()),
+                    model=model_name,
+                    output=[],  # Output should be populated by this point from text streaming
+                    status="completed",
+                    parallel_tool_calls=False,
+                    tool_choice="none",
+                    tools=[],
+                )
+
+                return [
+                    ResponseCompletedEvent(
+                        type="response.completed", sequence_number=self._next_sequence(context), response=response_obj
+                    )
+                ]
+
+            if event_class == "WorkflowFailedEvent":
+                workflow_id = context.get("workflow_id", str(uuid4()))
+                error_info = getattr(event, "error", None)
+
+                # Import Response and ResponseError types
+                from openai.types.responses import Response, ResponseError
+
+                # Get model name from context
+                model_name = context.get("request", {}).model if context.get("request") else "workflow"
+
+                # Create error object
+                error_message = str(error_info) if error_info else "Unknown error"
+
+                # Create ResponseError object (code must be one of the allowed values)
+                response_error = ResponseError(
+                    message=error_message,
+                    code="server_error",  # Use generic server_error code for workflow failures
+                )
+
+                # Create a full Response object for failed state
+                response_obj = Response(
+                    id=f"resp_{workflow_id}",
+                    object="response",
+                    created_at=float(time.time()),
+                    model=model_name,
+                    output=[],
+                    status="failed",
+                    error=response_error,
+                    parallel_tool_calls=False,
+                    tool_choice="none",
+                    tools=[],
+                )
+
+                return [
+                    ResponseFailedEvent(
+                        type="response.failed", sequence_number=self._next_sequence(context), response=response_obj
+                    )
+                ]
+
+            # Executor-level events (output items)
+            if event_class == "ExecutorInvokedEvent":
+                executor_id = getattr(event, "executor_id", "unknown")
+                item_id = f"exec_{executor_id}_{uuid4().hex[:8]}"
+                context[f"exec_item_{executor_id}"] = item_id
+                context["output_index"] = context.get("output_index", -1) + 1
+
+                # Create ExecutorActionItem with proper type
+                executor_item = ExecutorActionItem(
+                    type="executor_action",
+                    id=item_id,
+                    executor_id=executor_id,
+                    status="in_progress",
+                    metadata=getattr(event, "metadata", {}),
+                )
+
+                # Use our custom event type that accepts ExecutorActionItem
+                return [
+                    CustomResponseOutputItemAddedEvent(
+                        type="response.output_item.added",
+                        output_index=context["output_index"],
+                        sequence_number=self._next_sequence(context),
+                        item=executor_item,
+                    )
+                ]
+
+            if event_class == "ExecutorCompletedEvent":
+                executor_id = getattr(event, "executor_id", "unknown")
+                item_id = context.get(f"exec_item_{executor_id}", f"exec_{executor_id}_unknown")
+
+                # Create ExecutorActionItem with completed status
+                # ExecutorCompletedEvent uses 'data' field, not 'result'
+                executor_item = ExecutorActionItem(
+                    type="executor_action",
+                    id=item_id,
+                    executor_id=executor_id,
+                    status="completed",
+                    result=getattr(event, "data", None),
+                )
+
+                # Use our custom event type
+                return [
+                    CustomResponseOutputItemDoneEvent(
+                        type="response.output_item.done",
+                        output_index=context.get("output_index", 0),
+                        sequence_number=self._next_sequence(context),
+                        item=executor_item,
+                    )
+                ]
+
+            if event_class == "ExecutorFailedEvent":
+                executor_id = getattr(event, "executor_id", "unknown")
+                item_id = context.get(f"exec_item_{executor_id}", f"exec_{executor_id}_unknown")
+                error_info = getattr(event, "error", None)
+
+                # Create ExecutorActionItem with failed status
+                executor_item = ExecutorActionItem(
+                    type="executor_action",
+                    id=item_id,
+                    executor_id=executor_id,
+                    status="failed",
+                    error={"message": str(error_info)} if error_info else None,
+                )
+
+                # Use our custom event type
+                return [
+                    CustomResponseOutputItemDoneEvent(
+                        type="response.output_item.done",
+                        output_index=context.get("output_index", 0),
+                        sequence_number=self._next_sequence(context),
+                        item=executor_item,
+                    )
+                ]
+
+            # Handle informational workflow events (status, warnings, errors)
+            if event_class in ["WorkflowStatusEvent", "WorkflowWarningEvent", "WorkflowErrorEvent", "RequestInfoEvent"]:
+                # These are informational events that don't map to OpenAI lifecycle events
+                # Convert them to trace events for debugging visibility
+                event_data: dict[str, Any] = {}
+
+                # Extract relevant data based on event type
+                if event_class == "WorkflowStatusEvent":
+                    event_data["state"] = str(getattr(event, "state", "unknown"))
+                elif event_class == "WorkflowWarningEvent":
+                    event_data["message"] = str(getattr(event, "message", ""))
+                elif event_class == "WorkflowErrorEvent":
+                    event_data["message"] = str(getattr(event, "message", ""))
+                    event_data["error"] = str(getattr(event, "error", ""))
+                elif event_class == "RequestInfoEvent":
+                    request_info = getattr(event, "data", {})
+                    event_data["request_info"] = request_info if isinstance(request_info, dict) else str(request_info)
+
+                # Create a trace event for debugging
+                trace_event = ResponseTraceEventComplete(
+                    type="response.trace.complete",
+                    data={
+                        "trace_type": "workflow_info",
+                        "event_type": event_class,
+                        "data": event_data,
+                        "timestamp": datetime.now().isoformat(),
+                    },
+                    span_id=f"workflow_info_{uuid4().hex[:8]}",
+                    item_id=context["item_id"],
+                    output_index=context.get("output_index", 0),
+                    sequence_number=self._next_sequence(context),
+                )
+
+                return [trace_event]
+
+            # For unknown/legacy events, still emit as workflow event for backward compatibility
             # Get event data and serialize if it's a SerializationMixin
-            event_data = getattr(event, "data", None)
-            if event_data is not None and hasattr(event_data, "to_dict"):
+            raw_event_data = getattr(event, "data", None)
+            serialized_event_data: dict[str, Any] | str | None = raw_event_data
+            if raw_event_data is not None and hasattr(raw_event_data, "to_dict"):
                 # SerializationMixin objects - convert to dict for JSON serialization
                 try:
-                    event_data = event_data.to_dict()
+                    serialized_event_data = raw_event_data.to_dict()
                 except Exception as e:
                     logger.debug(f"Failed to serialize event data with to_dict(): {e}")
-                    event_data = str(event_data)
+                    serialized_event_data = str(raw_event_data)
 
-            # Create structured workflow event
+            # Create structured workflow event (keeping for backward compatibility)
             workflow_event = ResponseWorkflowEventComplete(
                 type="response.workflow_event.complete",
                 data={
                     "event_type": event.__class__.__name__,
-                    "data": event_data,
+                    "data": serialized_event_data,
                     "executor_id": getattr(event, "executor_id", None),
                     "timestamp": datetime.now().isoformat(),
                 },
@@ -394,6 +840,7 @@ class MessageMapper:
                 sequence_number=self._next_sequence(context),
             )
 
+            logger.debug(f"Unhandled workflow event type: {event_class}, emitting as legacy workflow event")
             return [workflow_event]
 
         except Exception as e:
@@ -538,8 +985,16 @@ class MessageMapper:
         result = getattr(content, "result", None)
         exception = getattr(content, "exception", None)
 
-        # Convert result to string
-        output = result if isinstance(result, str) else json.dumps(result) if result is not None else ""
+        # Convert result to string, handling nested Content objects from MCP tools
+        if isinstance(result, str):
+            output = result
+        elif result is not None:
+            # Recursively serialize any nested Content objects (e.g., from MCP tools)
+            serialized = _serialize_content_recursive(result)
+            # Convert to JSON string if still not a string
+            output = serialized if isinstance(serialized, str) else json.dumps(serialized)
+        else:
+            output = ""
 
         # Determine status based on exception
         status = "incomplete" if exception else "completed"
@@ -556,6 +1011,7 @@ class MessageMapper:
             item_id=item_id,
             output_index=context["output_index"],
             sequence_number=self._next_sequence(context),
+            timestamp=datetime.now().isoformat(),
         )
 
     async def _map_error_content(self, content: Any, context: dict[str, Any]) -> ResponseErrorEvent:
@@ -723,7 +1179,7 @@ class MessageMapper:
     async def _create_unknown_content_event(self, content: Any, context: dict[str, Any]) -> ResponseStreamEvent:
         """Create event for unknown content types."""
         content_type = content.__class__.__name__
-        text = f"⚠️ Unknown content type: {content_type}\n"
+        text = f"Warning: Unknown content type: {content_type}\n"
         return self._create_text_delta_event(text, context)
 
     async def _create_error_response(self, error_message: str, request: AgentFrameworkRequest) -> OpenAIResponse:

python/packages/devui/agent_framework_devui/_server.py

@@ -85,19 +85,25 @@ class DevServer:
         return self.executor
 
     async def _cleanup_entities(self) -> None:
-        """Cleanup entity resources (close clients, credentials, etc.)."""
+        """Cleanup entity resources (close clients, MCP tools, credentials, etc.)."""
         if not self.executor:
             return
 
         logger.info("Cleaning up entity resources...")
         entities = self.executor.entity_discovery.list_entities()
         closed_count = 0
+        mcp_tools_closed = 0
+        credentials_closed = 0
 
         for entity_info in entities:
             try:
                 entity_obj = self.executor.entity_discovery.get_entity_object(entity_info.id)
+
+                # Close chat clients and their credentials
                 if entity_obj and hasattr(entity_obj, "chat_client"):
                     client = entity_obj.chat_client
+
+                    # Close the chat client itself
                     if hasattr(client, "close") and callable(client.close):
                         if inspect.iscoroutinefunction(client.close):
                             await client.close()
@@ -105,11 +111,47 @@ class DevServer:
                             client.close()
                         closed_count += 1
                         logger.debug(f"Closed client for entity: {entity_info.id}")
+
+                    # Close credentials attached to chat clients (e.g., AzureCliCredential)
+                    credential_attrs = ["credential", "async_credential", "_credential", "_async_credential"]
+                    for attr in credential_attrs:
+                        if hasattr(client, attr):
+                            cred = getattr(client, attr)
+                            if cred and hasattr(cred, "close") and callable(cred.close):
+                                try:
+                                    if inspect.iscoroutinefunction(cred.close):
+                                        await cred.close()
+                                    else:
+                                        cred.close()
+                                    credentials_closed += 1
+                                    logger.debug(f"Closed credential for entity: {entity_info.id}")
+                                except Exception as e:
+                                    logger.warning(f"Error closing credential for {entity_info.id}: {e}")
+
+                # Close MCP tools (framework tracks them in _local_mcp_tools)
+                if entity_obj and hasattr(entity_obj, "_local_mcp_tools"):
+                    for mcp_tool in entity_obj._local_mcp_tools:
+                        if hasattr(mcp_tool, "close") and callable(mcp_tool.close):
+                            try:
+                                if inspect.iscoroutinefunction(mcp_tool.close):
+                                    await mcp_tool.close()
+                                else:
+                                    mcp_tool.close()
+                                mcp_tools_closed += 1
+                                tool_name = getattr(mcp_tool, "name", "unknown")
+                                logger.debug(f"Closed MCP tool '{tool_name}' for entity: {entity_info.id}")
+                            except Exception as e:
+                                logger.warning(f"Error closing MCP tool for {entity_info.id}: {e}")
+
             except Exception as e:
                 logger.warning(f"Error closing entity {entity_info.id}: {e}")
 
         if closed_count > 0:
             logger.info(f"Closed {closed_count} entity client(s)")
+        if credentials_closed > 0:
+            logger.info(f"Closed {credentials_closed} credential(s)")
+        if mcp_tools_closed > 0:
+            logger.info(f"Closed {mcp_tools_closed} MCP tool(s)")
 
     def create_app(self) -> FastAPI:
         """Create the FastAPI application."""

python/packages/devui/agent_framework_devui/models/__init__.py

@@ -30,6 +30,9 @@ from openai.types.shared import Metadata, ResponsesModel
 from ._discovery_models import DiscoveryResponse, EntityInfo
 from ._openai_custom import (
     AgentFrameworkRequest,
+    CustomResponseOutputItemAddedEvent,
+    CustomResponseOutputItemDoneEvent,
+    ExecutorActionItem,
     OpenAIError,
     ResponseFunctionResultComplete,
     ResponseTraceEvent,
@@ -46,8 +49,11 @@ __all__ = [
     "Conversation",
     "ConversationDeletedResource",
     "ConversationItem",
+    "CustomResponseOutputItemAddedEvent",
+    "CustomResponseOutputItemDoneEvent",
     "DiscoveryResponse",
     "EntityInfo",
+    "ExecutorActionItem",
     "InputTokensDetails",
     "Metadata",
     "OpenAIError",

python/packages/devui/agent_framework_devui/models/_openai_custom.py

@@ -8,6 +8,7 @@ to support Agent Framework specific features like workflows and traces.
 
 from __future__ import annotations
 
+from dataclasses import dataclass
 from typing import Any, Literal
 
 from pydantic import BaseModel, ConfigDict
@@ -15,6 +16,69 @@ from pydantic import BaseModel, ConfigDict
 # Custom Agent Framework OpenAI event types for structured data
 
 
+# Agent lifecycle events - simple and clear
+class AgentStartedEvent:
+    """Event emitted when an agent starts execution."""
+
+    pass
+
+
+class AgentCompletedEvent:
+    """Event emitted when an agent completes execution successfully."""
+
+    pass
+
+
+@dataclass
+class AgentFailedEvent:
+    """Event emitted when an agent fails during execution."""
+
+    error: Exception | None = None
+
+
+class ExecutorActionItem(BaseModel):
+    """Custom item type for workflow executor actions.
+
+    This is a DevUI-specific extension to represent workflow executors as output items.
+    Since OpenAI's ResponseOutputItemAddedEvent only accepts specific item types,
+    and executor actions are not part of the standard, we need this custom type.
+    """
+
+    type: Literal["executor_action"] = "executor_action"
+    id: str
+    executor_id: str
+    status: Literal["in_progress", "completed", "failed", "cancelled"] = "in_progress"
+    metadata: dict[str, Any] | None = None
+    result: Any | None = None
+    error: dict[str, Any] | None = None
+
+
+class CustomResponseOutputItemAddedEvent(BaseModel):
+    """Custom version of ResponseOutputItemAddedEvent that accepts any item type.
+
+    This allows us to emit executor action items while maintaining the same
+    event structure as OpenAI's standard.
+    """
+
+    type: Literal["response.output_item.added"] = "response.output_item.added"
+    output_index: int
+    sequence_number: int
+    item: dict[str, Any] | ExecutorActionItem | Any  # Flexible item type
+
+
+class CustomResponseOutputItemDoneEvent(BaseModel):
+    """Custom version of ResponseOutputItemDoneEvent that accepts any item type.
+
+    This allows us to emit executor action items while maintaining the same
+    event structure as OpenAI's standard.
+    """
+
+    type: Literal["response.output_item.done"] = "response.output_item.done"
+    output_index: int
+    sequence_number: int
+    item: dict[str, Any] | ExecutorActionItem | Any  # Flexible item type
+
+
 class ResponseWorkflowEventComplete(BaseModel):
     """Complete workflow event data."""
 
@@ -57,6 +121,7 @@ class ResponseFunctionResultComplete(BaseModel):
     item_id: str
     output_index: int = 0
     sequence_number: int
+    timestamp: str | None = None  # Optional timestamp for UI display
 
 
 # Agent Framework extension fields
@@ -64,7 +129,7 @@ class AgentFrameworkExtraBody(BaseModel):
     """Agent Framework specific routing fields for OpenAI requests."""
 
     entity_id: str
-    input_data: dict[str, Any] | None = None
+    # input_data removed - now using standard input field for all data
 
     model_config = ConfigDict(extra="allow")
 
@@ -80,7 +145,7 @@ class AgentFrameworkRequest(BaseModel):
 
     # All OpenAI fields from ResponseCreateParams
     model: str  # Used as entity_id in DevUI!
-    input: str | list[Any]  # ResponseInputParam
+    input: str | list[Any] | dict[str, Any]  # ResponseInputParam + dict for workflow structured input
     stream: bool | None = False
 
     # OpenAI conversation parameter (standard!)

python/packages/devui/agent_framework_devui/ui/index.html

@@ -5,7 +5,7 @@
     <link rel="icon" type="image/svg+xml" href="/agentframework.svg" />
     <meta name="viewport" content="width=device-width, initial-scale=1.0" />
     <title>Agent Framework Dev UI</title>
-    <script type="module" crossorigin src="/assets/index-DmL7WSFa.js"></script>
+    <script type="module" crossorigin src="/assets/index-D_Y1oSGu.js"></script>
     <link rel="stylesheet" crossorigin href="/assets/index-CE4pGoXh.css">
   </head>
   <body>