resolved the human in the loop issue

Author: Dhruvkumar-Microsoft

src/backend/v4/magentic_agents/proxy_agent.py

@@ -147,16 +147,19 @@ async def _invoke_stream_internal(
         logger.debug("ProxyAgent: Message text: %s", message_text[:100])
 
         clarification_req_text = f"{message_text}"
+        request_id = str(uuid.uuid4())
         clarification_request = UserClarificationRequest(
             question=clarification_req_text,
-            request_id=str(uuid.uuid4()),
+            request_id=request_id,
         )
 
         # Dispatch websocket event requesting clarification
+        # Serialize dataclass to a plain dict so json.dumps produces proper JSON
+        # instead of relying on str() repr which is fragile for the frontend parser.
         await connection_config.send_status_update_async(
             {
-                "type": WebsocketMessageType.USER_CLARIFICATION_REQUEST,
-                "data": clarification_request,
+                "question": clarification_req_text,
+                "request_id": request_id,
             },
             user_id=self.user_id,
             message_type=WebsocketMessageType.USER_CLARIFICATION_REQUEST,

src/backend/v4/orchestration/human_approval_manager.py

@@ -8,7 +8,7 @@
 from typing import Any, Optional
 
 import v4.models.messages as messages
-from agent_framework import Message
+from agent_framework import AgentResponse, Message
 from agent_framework_orchestrations._magentic import (
     MagenticContext,
     StandardMagenticManager,
@@ -94,6 +94,45 @@ def __init__(self, user_id: str, agent, *args, **kwargs):
         # New API: StandardMagenticManager takes agent as first positional argument
         super().__init__(agent, *args, **kwargs)
 
+    async def _complete(self, messages: list[Message]) -> Message:
+        """Override to pass session=None, making each LLM call stateless.
+
+        The base class passes session=self._session which triggers
+        InMemoryHistoryProvider auto-injection and previous_response_id
+        chaining in rc4. This causes message payloads to grow with every
+        internal call (facts, plan, progress ledger, etc.), burning through
+        TPM quota (429 errors) and confusing the orchestrator LLM's routing
+        decisions (e.g. skipping ProxyAgent for user clarification).
+
+        Passing session=None restores the old stateless behavior where each
+        call only sends the messages explicitly provided.
+        """
+        from openai import RateLimitError
+
+        max_retries = 5
+        base_delay = 2.0  # seconds
+
+        for attempt in range(max_retries):
+            try:
+                response: AgentResponse = await self._agent.run(messages, session=None)
+                if not response.messages:
+                    raise RuntimeError("Agent returned no messages in response.")
+                if len(response.messages) > 1:
+                    logger.warning("Agent returned multiple messages; using the last one.")
+                return response.messages[-1]
+            except Exception as exc:
+                inner = getattr(exc, "inner_exception", None)
+                is_rate_limit = isinstance(inner, RateLimitError) or "429" in str(exc)
+                if is_rate_limit and attempt < max_retries - 1:
+                    delay = base_delay * (2 ** attempt)
+                    logger.warning(
+                        "Rate limit hit (attempt %d/%d). Retrying in %.1fs...",
+                        attempt + 1, max_retries, delay,
+                    )
+                    await asyncio.sleep(delay)
+                    continue
+                raise
+
     async def plan(self, magentic_context: MagenticContext) -> Any:
         """
         Override the plan method to create the plan first, then ask for approval before execution.

src/backend/v4/orchestration/orchestration_manager.py

@@ -14,15 +14,13 @@
     ChatOptions,
     Message,
     InMemoryCheckpointStorage,
-    WorkflowEvent,
 )
 from agent_framework_orchestrations import MagenticBuilder
 from agent_framework_orchestrations._base_group_chat_orchestrator import (
     GroupChatRequestSentEvent,
     GroupChatResponseReceivedEvent,
 )
 from agent_framework_orchestrations._magentic import (
-    MagenticOrchestratorEvent,
     MagenticProgressLedger,
 )
 
@@ -34,7 +32,6 @@
 from v4.common.services.team_service import TeamService
 import time as _time
 from v4.callbacks.response_handlers import (
-    agent_response_callback,
     streaming_agent_response_callback,
 )
 from v4.config.settings import connection_config, orchestration_config
@@ -387,7 +384,6 @@ async def run_orchestration(self, user_id: str, input_task) -> None:
 
             self.logger.info("Starting workflow execution...")
 
-            last_message_id: str | None = None
             async for event in workflow.run(task_text, stream=True):
                 try:
                     # WorkflowEvent has a .type field (string) instead of specific event classes

src/frontend/src/services/PlanDataService.tsx

@@ -765,6 +765,23 @@ export class PlanDataService {
    */
   static parseUserClarificationRequest(rawData: any): ParsedUserClarification | null {
     try {
+      // First try direct JSON extraction (clean dict format from backend)
+      const extractDirect = (val: any, depth = 0): ParsedUserClarification | null => {
+        if (depth > 10 || !val || typeof val !== 'object') return null;
+        if (typeof val.question === 'string' && typeof val.request_id === 'string') {
+          return {
+            type: WebsocketMessageType.USER_CLARIFICATION_REQUEST,
+            question: val.question.trim(),
+            request_id: val.request_id,
+          };
+        }
+        if (val.data !== undefined) return extractDirect(val.data, depth + 1);
+        return null;
+      };
+      const direct = extractDirect(rawData);
+      if (direct) return direct;
+
+      // Fallback: extract from Python repr string (legacy format)
       const extractString = (val: any, depth = 0): string | null => {
         if (depth > 15) return null;
         if (typeof val === 'string') {