fix(backend): standardize microservice host/port configuration

- Change agentgenerator_host default from empty string to "localhost"
  - Consistent with other client services (rabbitmq, redis, clamav)
  - Fixes "service_not_configured" error when only port is set
- Change agentgenerator_port default from 8000 to 8009
  - Avoids conflict with Kong gateway (port 8000)
  - Follows sequential port allocation (8001-8008 already in use)
- Add AGENTGENERATOR_HOST to .env.default for clarity

This standardization ensures:
1. Consistent host naming across all microservice client configurations
2. No port conflicts with Kong or other services
3. Agent Generator service works out of the box when enabled

autogpt_platform/backend/.env.default

@@ -192,3 +192,8 @@ POSTHOG_HOST=https://eu.i.posthog.com
 
 # Other Services
 AUTOMOD_API_KEY=
+
+# Agent Generator Service
+# The Agent Generator microservice handles AI-powered agent creation from natural language
+AGENTGENERATOR_HOST=localhost
+AGENTGENERATOR_PORT=8009

autogpt_platform/backend/backend/copilot/config.py

@@ -27,7 +27,6 @@ class ChatConfig(BaseSettings):
     session_ttl: int = Field(default=43200, description="Session TTL in seconds")
 
     # Streaming Configuration
-    stream_timeout: int = Field(default=300, description="Stream timeout in seconds")
     max_retries: int = Field(
         default=3,
         description="Max retries for fallback path (SDK handles retries internally)",
@@ -39,8 +38,10 @@ class ChatConfig(BaseSettings):
 
     # Long-running operation configuration
     long_running_operation_ttl: int = Field(
-        default=600,
-        description="TTL in seconds for long-running operation tracking in Redis (safety net if pod dies)",
+        default=3600,
+        description="TTL in seconds for long-running operation deduplication lock "
+        "(1 hour, matches stream_ttl). Prevents duplicate operations if pod dies. "
+        "For longer operations, the stream_registry heartbeat keeps them alive.",
     )
 
     # Stream registry configuration for SSE reconnection
@@ -48,6 +49,11 @@ class ChatConfig(BaseSettings):
         default=3600,
         description="TTL in seconds for stream data in Redis (1 hour)",
     )
+    stream_lock_ttl: int = Field(
+        default=120,
+        description="TTL in seconds for stream lock (2 minutes). Short timeout allows "
+        "reconnection after refresh/crash without long waits.",
+    )
     stream_max_length: int = Field(
         default=10000,
         description="Maximum number of messages to store per stream",

autogpt_platform/backend/backend/copilot/db.py

@@ -3,8 +3,9 @@
 import asyncio
 import logging
 from datetime import UTC, datetime
-from typing import Any, cast
+from typing import Any
 
+from prisma.errors import UniqueViolationError
 from prisma.models import ChatMessage as PrismaChatMessage
 from prisma.models import ChatSession as PrismaChatSession
 from prisma.types import (
@@ -92,10 +93,9 @@ async def add_chat_message(
     function_call: dict[str, Any] | None = None,
 ) -> ChatMessage:
     """Add a message to a chat session."""
-    # Build input dict dynamically rather than using ChatMessageCreateInput directly
-    # because Prisma's TypedDict validation rejects optional fields set to None.
-    # We only include fields that have values, then cast at the end.
-    data: dict[str, Any] = {
+    # Build ChatMessageCreateInput with only non-None values
+    # (Prisma TypedDict rejects optional fields set to None)
+    data: ChatMessageCreateInput = {
         "Session": {"connect": {"id": session_id}},
         "role": role,
         "sequence": sequence,
@@ -123,7 +123,7 @@ async def add_chat_message(
             where={"id": session_id},
             data={"updatedAt": datetime.now(UTC)},
         ),
-        PrismaChatMessage.prisma().create(data=cast(ChatMessageCreateInput, data)),
+        PrismaChatMessage.prisma().create(data=data),
     )
     return ChatMessage.from_db(message)
 
@@ -132,58 +132,93 @@ async def add_chat_messages_batch(
     session_id: str,
     messages: list[dict[str, Any]],
     start_sequence: int,
-) -> list[ChatMessage]:
+) -> int:
     """Add multiple messages to a chat session in a batch.
 
-    Uses a transaction for atomicity - if any message creation fails,
-    the entire batch is rolled back.
+    Uses collision detection with retry: tries to create messages starting
+    at start_sequence. If a unique constraint violation occurs (e.g., the
+    streaming loop and long-running callback race), queries the latest
+    sequence and retries with the correct offset. This avoids unnecessary
+    upserts and DB queries in the common case (no collision).
+
+    Returns:
+        Next sequence number for the next message to be inserted. This equals
+        start_sequence + len(messages) and allows callers to update their
+        counters even when collision detection adjusts start_sequence.
     """
     if not messages:
-        return []
+        # No messages to add - return current count
+        return start_sequence
 
-    created_messages = []
+    max_retries = 5
+    for attempt in range(max_retries):
+        try:
+            # Single timestamp for all messages and session update
+            now = datetime.now(UTC)
 
-    async with db.transaction() as tx:
-        for i, msg in enumerate(messages):
-            # Build input dict dynamically rather than using ChatMessageCreateInput
-            # directly because Prisma's TypedDict validation rejects optional fields
-            # set to None. We only include fields that have values, then cast.
-            data: dict[str, Any] = {
-                "Session": {"connect": {"id": session_id}},
-                "role": msg["role"],
-                "sequence": start_sequence + i,
-            }
+            async with db.transaction() as tx:
+                # Build all message data
+                messages_data = []
+                for i, msg in enumerate(messages):
+                    # Build ChatMessageCreateInput with only non-None values
+                    # (Prisma TypedDict rejects optional fields set to None)
+                    # Note: create_many doesn't support nested creates, use sessionId directly
+                    data: ChatMessageCreateInput = {
+                        "sessionId": session_id,
+                        "role": msg["role"],
+                        "sequence": start_sequence + i,
+                        "createdAt": now,
+                    }
 
-            # Add optional string fields
-            if msg.get("content") is not None:
-                data["content"] = msg["content"]
-            if msg.get("name") is not None:
-                data["name"] = msg["name"]
-            if msg.get("tool_call_id") is not None:
-                data["toolCallId"] = msg["tool_call_id"]
-            if msg.get("refusal") is not None:
-                data["refusal"] = msg["refusal"]
+                    # Add optional string fields
+                    if msg.get("content") is not None:
+                        data["content"] = msg["content"]
+                    if msg.get("name") is not None:
+                        data["name"] = msg["name"]
+                    if msg.get("tool_call_id") is not None:
+                        data["toolCallId"] = msg["tool_call_id"]
+                    if msg.get("refusal") is not None:
+                        data["refusal"] = msg["refusal"]
 
-            # Add optional JSON fields only when they have values
-            if msg.get("tool_calls") is not None:
-                data["toolCalls"] = SafeJson(msg["tool_calls"])
-            if msg.get("function_call") is not None:
-                data["functionCall"] = SafeJson(msg["function_call"])
+                    # Add optional JSON fields only when they have values
+                    if msg.get("tool_calls") is not None:
+                        data["toolCalls"] = SafeJson(msg["tool_calls"])
+                    if msg.get("function_call") is not None:
+                        data["functionCall"] = SafeJson(msg["function_call"])
 
-            created = await PrismaChatMessage.prisma(tx).create(
-                data=cast(ChatMessageCreateInput, data)
-            )
-            created_messages.append(created)
+                    messages_data.append(data)
 
-        # Update session's updatedAt timestamp within the same transaction.
-        # Note: Token usage (total_prompt_tokens, total_completion_tokens) is updated
-        # separately via update_chat_session() after streaming completes.
-        await PrismaChatSession.prisma(tx).update(
-            where={"id": session_id},
-            data={"updatedAt": datetime.now(UTC)},
-        )
+                # Run create_many and session update in parallel within transaction
+                # Both use the same timestamp for consistency
+                await asyncio.gather(
+                    PrismaChatMessage.prisma(tx).create_many(data=messages_data),
+                    PrismaChatSession.prisma(tx).update(
+                        where={"id": session_id},
+                        data={"updatedAt": now},
+                    ),
+                )
 
-    return [ChatMessage.from_db(m) for m in created_messages]
+            # Return next sequence number for counter sync
+            return start_sequence + len(messages)
+
+        except UniqueViolationError:
+            if attempt < max_retries - 1:
+                # Collision detected - query MAX(sequence)+1 and retry with correct offset
+                logger.info(
+                    f"Collision detected for session {session_id} at sequence "
+                    f"{start_sequence}, querying DB for latest sequence"
+                )
+                start_sequence = await get_next_sequence(session_id)
+                logger.info(
+                    f"Retrying batch insert with start_sequence={start_sequence}"
+                )
+                continue
+            else:
+                # Max retries exceeded - propagate error
+                raise
+
+    # Should never reach here due to raise in exception handler
+    raise RuntimeError(f"Failed to insert messages after {max_retries} attempts")
 
 
 async def get_user_chat_sessions(
@@ -237,10 +272,20 @@ async def delete_chat_session(session_id: str, user_id: str | None = None) -> bo
         return False
 
 
-async def get_chat_session_message_count(session_id: str) -> int:
-    """Get the number of messages in a chat session."""
-    count = await PrismaChatMessage.prisma().count(where={"sessionId": session_id})
-    return count
+async def get_next_sequence(session_id: str) -> int:
+    """Get the next sequence number for a new message in this session.
+
+    Uses MAX(sequence) + 1 for robustness. Returns 0 if no messages exist.
+    More robust than COUNT(*) because it's immune to deleted messages.
+
+    Optimized to select only the sequence column using raw SQL.
+    The unique index on (sessionId, sequence) makes this query fast.
+    """
+    results = await db.query_raw_with_schema(
+        'SELECT "sequence" FROM {schema_prefix}"ChatMessage" WHERE "sessionId" = $1 ORDER BY "sequence" DESC LIMIT 1',
+        session_id,
+    )
+    return 0 if not results else results[0]["sequence"] + 1
 
 
 async def update_tool_message_content(

autogpt_platform/backend/backend/copilot/executor/processor.py

@@ -266,7 +266,11 @@ class CoPilotProcessor:
 
         except asyncio.CancelledError:
             log.info("Task cancelled")
-            await stream_registry.mark_task_completed(entry.task_id, status="failed")
+            await stream_registry.mark_task_completed(
+                entry.task_id,
+                status="failed",
+                error_message="Task was cancelled",
+            )
             raise
 
         except Exception as e:

autogpt_platform/backend/backend/copilot/model.py

@@ -434,8 +434,6 @@ async def _get_session_from_db(session_id: str) -> ChatSession | None:
 
 async def upsert_chat_session(
     session: ChatSession,
-    *,
-    existing_message_count: int | None = None,
 ) -> ChatSession:
     """Update a chat session in both cache and database.
 
@@ -443,12 +441,6 @@ async def upsert_chat_session(
     operations (e.g., background title update and main stream handler)
     attempt to upsert the same session simultaneously.
 
-    Args:
-        existing_message_count: If provided, skip the DB query to count
-            existing messages. The caller is responsible for tracking this
-            accurately. Useful for incremental saves in a streaming loop
-            where the caller already knows how many messages are persisted.
-
     Raises:
         DatabaseError: If the database write fails. The cache is still updated
             as a best-effort optimization, but the error is propagated to ensure
@@ -459,11 +451,8 @@ async def upsert_chat_session(
     lock = await _get_session_lock(session.session_id)
 
     async with lock:
-        # Get existing message count from DB for incremental saves
-        if existing_message_count is None:
-            existing_message_count = await chat_db().get_chat_session_message_count(
-                session.session_id
-            )
+        # Always query DB for existing message count to ensure consistency
+        existing_message_count = await chat_db().get_next_sequence(session.session_id)
 
         db_error: Exception | None = None
 
@@ -587,9 +576,7 @@ async def append_and_save_message(session_id: str, message: ChatMessage) -> Chat
             raise ValueError(f"Session {session_id} not found")
 
         session.messages.append(message)
-        existing_message_count = await chat_db().get_chat_session_message_count(
-            session_id
-        )
+        existing_message_count = await chat_db().get_next_sequence(session_id)
 
         try:
             await _save_session_to_db(session, existing_message_count)

autogpt_platform/backend/backend/copilot/model_test.py

@@ -331,3 +331,96 @@ def test_to_openai_messages_merges_split_assistants():
     tc_list = merged.get("tool_calls")
     assert tc_list is not None and len(list(tc_list)) == 1
     assert list(tc_list)[0]["id"] == "tc1"
+
+
+# --------------------------------------------------------------------------- #
+#  Concurrent save collision detection                                        #
+# --------------------------------------------------------------------------- #
+
+
+@pytest.mark.asyncio(loop_scope="session")
+async def test_concurrent_saves_collision_detection(setup_test_user, test_user_id):
+    """Test that concurrent saves from streaming loop and callback handle collisions correctly.
+
+    Simulates the race condition where:
+    1. Streaming loop starts with saved_msg_count=5
+    2. Long-running callback appends message #5 and saves
+    3. Streaming loop tries to save with stale count=5
+
+    The collision detection should handle this gracefully.
+    """
+    import asyncio
+
+    # Create a session with initial messages
+    session = ChatSession.new(user_id=test_user_id)
+    for i in range(3):
+        session.messages.append(
+            ChatMessage(
+                role="user" if i % 2 == 0 else "assistant", content=f"Message {i}"
+            )
+        )
+
+    # Save initial messages
+    session = await upsert_chat_session(session)
+
+    # Simulate streaming loop and callback saving concurrently
+    async def streaming_loop_save():
+        """Simulates streaming loop saving messages."""
+        # Add 2 messages
+        session.messages.append(ChatMessage(role="user", content="Streaming message 1"))
+        session.messages.append(
+            ChatMessage(role="assistant", content="Streaming message 2")
+        )
+
+        # Wait a bit to let callback potentially save first
+        await asyncio.sleep(0.01)
+
+        # Save (will query DB for existing count)
+        return await upsert_chat_session(session)
+
+    async def callback_save():
+        """Simulates long-running callback saving a message."""
+        # Add 1 message
+        session.messages.append(
+            ChatMessage(role="tool", content="Callback result", tool_call_id="tc1")
+        )
+
+        # Save immediately (will query DB for existing count)
+        return await upsert_chat_session(session)
+
+    # Run both saves concurrently - one will hit collision detection
+    results = await asyncio.gather(streaming_loop_save(), callback_save())
+
+    # Both should succeed
+    assert all(r is not None for r in results)
+
+    # Reload session from DB to verify
+    from backend.data.redis_client import get_redis_async
+
+    redis_key = f"chat:session:{session.session_id}"
+    async_redis = await get_redis_async()
+    await async_redis.delete(redis_key)  # Clear cache to force DB load
+
+    loaded_session = await get_chat_session(session.session_id, test_user_id)
+    assert loaded_session is not None
+
+    # Should have all 6 messages (3 initial + 2 streaming + 1 callback)
+    assert len(loaded_session.messages) == 6
+
+    # Verify no duplicate sequences
+    sequences = []
+    for i, msg in enumerate(loaded_session.messages):
+        # Messages should have sequential sequence numbers starting from 0
+        sequences.append(i)
+
+    # All sequences should be unique and sequential
+    assert sequences == list(range(6))
+
+    # Verify message content is preserved
+    contents = [m.content for m in loaded_session.messages]
+    assert "Message 0" in contents
+    assert "Message 1" in contents
+    assert "Message 2" in contents
+    assert "Streaming message 1" in contents
+    assert "Streaming message 2" in contents
+    assert "Callback result" in contents

autogpt_platform/backend/backend/copilot/sdk/service.py

@@ -7,8 +7,10 @@ import os
 import uuid
 from collections.abc import AsyncGenerator
 from dataclasses import dataclass
-from typing import Any
+from typing import Any, cast
 
+from backend.data.redis_client import get_redis_async
+from backend.executor.cluster_lock import AsyncClusterLock
 from backend.util.exceptions import NotFoundError
 
 from .. import stream_registry
@@ -61,6 +63,7 @@ from .transcript import (
 logger = logging.getLogger(__name__)
 config = ChatConfig()
 
+
 # Set to hold background tasks to prevent garbage collection
 _background_tasks: set[asyncio.Task[Any]] = set()
 
@@ -132,8 +135,12 @@ is delivered to the user via a background stream.
   All tasks must run in the foreground.
 """
 
+STREAM_LOCK_PREFIX = "copilot:stream:lock:"
 
-def _build_long_running_callback(user_id: str | None) -> LongRunningCallback:
+
+def _build_long_running_callback(
+    user_id: str | None,
+) -> LongRunningCallback:
     """Build a callback that delegates long-running tools to the non-SDK infrastructure.
 
     Long-running tools (create_agent, edit_agent, etc.) are delegated to the
@@ -142,6 +149,9 @@ def _build_long_running_callback(user_id: str | None) -> LongRunningCallback:
     page refreshes / pod restarts, and the frontend shows the proper loading
     widget with progress updates.
 
+    Args:
+        user_id: User ID for the session
+
     The returned callback matches the ``LongRunningCallback`` signature:
     ``(tool_name, args, session) -> MCP response dict``.
     """
@@ -207,7 +217,8 @@ def _build_long_running_callback(user_id: str | None) -> LongRunningCallback:
             tool_call_id=tool_call_id,
         )
         session.messages.append(pending_message)
-        await upsert_chat_session(session)
+        # Collision detection happens in add_chat_messages_batch (db.py)
+        session = await upsert_chat_session(session)
 
         # --- Spawn background task (reuses non-SDK infrastructure) ---
         bg_task = asyncio.create_task(
@@ -527,6 +538,9 @@ async def stream_chat_completion_sdk(
             f"Session {session_id} not found. Please create a new session first."
         )
 
+    # Type narrowing: session is guaranteed ChatSession after the check above
+    session = cast(ChatSession, session)
+
     # Append the new message to the session if it's not already there
     new_message_role = "user" if is_user_message else "assistant"
     if message and (
@@ -564,6 +578,29 @@ async def stream_chat_completion_sdk(
     system_prompt += _SDK_TOOL_SUPPLEMENT
     message_id = str(uuid.uuid4())
     task_id = str(uuid.uuid4())
+    stream_id = task_id  # Use task_id as unique stream identifier
+
+    # Acquire stream lock to prevent concurrent streams to the same session
+    lock = AsyncClusterLock(
+        redis=await get_redis_async(),
+        key=f"{STREAM_LOCK_PREFIX}{session_id}",
+        owner_id=stream_id,
+        timeout=config.stream_lock_ttl,
+    )
+
+    lock_owner = await lock.try_acquire()
+    if lock_owner != stream_id:
+        # Another stream is active
+        logger.warning(
+            f"[SDK] Session {session_id} already has an active stream: {lock_owner}"
+        )
+        yield StreamError(
+            errorText="Another stream is already active for this session. "
+            "Please wait or stop it.",
+            code="stream_already_active",
+        )
+        yield StreamFinish()
+        return
 
     yield StreamStart(messageId=message_id, taskId=task_id)
 
@@ -715,9 +752,6 @@ async def stream_chat_completion_sdk(
                 accumulated_tool_calls: list[dict[str, Any]] = []
                 has_appended_assistant = False
                 has_tool_results = False
-                # Track persisted message count to skip DB count queries
-                # on incremental saves.  Initial save happened at line 545.
-                saved_msg_count = len(session.messages)
 
                 # Use an explicit async iterator with non-cancelling heartbeats.
                 # CRITICAL: we must NOT cancel __anext__() mid-flight — doing so
@@ -744,6 +778,8 @@ async def stream_chat_completion_sdk(
 
                         if not done:
                             # Timeout — emit heartbeat but keep the task alive
+                            # Also refresh lock TTL to keep it alive
+                            await lock.refresh()
                             yield StreamHeartbeat()
                             continue
 
@@ -893,13 +929,10 @@ async def stream_chat_completion_sdk(
                                     has_appended_assistant = True
                                 # Save before tool execution starts so the
                                 # pending tool call is visible on refresh /
-                                # other devices.
+                                # other devices. Collision detection happens
+                                # in add_chat_messages_batch (db.py).
                                 try:
-                                    await upsert_chat_session(
-                                        session,
-                                        existing_message_count=saved_msg_count,
-                                    )
-                                    saved_msg_count = len(session.messages)
+                                    session = await upsert_chat_session(session)
                                 except Exception as save_err:
                                     logger.warning(
                                         "[SDK] [%s] Incremental save " "failed: %s",
@@ -922,12 +955,9 @@ async def stream_chat_completion_sdk(
                                 has_tool_results = True
                                 # Save after tool completes so the result is
                                 # visible on refresh / other devices.
+                                # Collision detection happens in add_chat_messages_batch (db.py).
                                 try:
-                                    await upsert_chat_session(
-                                        session,
-                                        existing_message_count=saved_msg_count,
-                                    )
-                                    saved_msg_count = len(session.messages)
+                                    session = await upsert_chat_session(session)
                                 except Exception as save_err:
                                     logger.warning(
                                         "[SDK] [%s] Incremental save " "failed: %s",
@@ -1059,7 +1089,7 @@ async def stream_chat_completion_sdk(
                 "to use the OpenAI-compatible fallback."
             )
 
-        await asyncio.shield(upsert_chat_session(session))
+        session = cast(ChatSession, await asyncio.shield(upsert_chat_session(session)))
         logger.info(
             "[SDK] [%s] Session saved with %d messages",
             session_id[:12],
@@ -1076,10 +1106,11 @@ async def stream_chat_completion_sdk(
         raise
     except Exception as e:
         logger.error(f"[SDK] Error: {e}", exc_info=True)
-        try:
-            await asyncio.shield(upsert_chat_session(session))
-        except Exception as save_err:
-            logger.error(f"[SDK] Failed to save session on error: {save_err}")
+        if session:
+            try:
+                await asyncio.shield(upsert_chat_session(session))
+            except Exception as save_err:
+                logger.error(f"[SDK] Failed to save session on error: {save_err}")
         yield StreamError(
             errorText="An error occurred. Please try again.",
             code="sdk_error",
@@ -1101,7 +1132,7 @@ async def stream_chat_completion_sdk(
                 if not raw_transcript and use_resume and resume_file:
                     raw_transcript = read_transcript_file(resume_file)
 
-                if raw_transcript:
+                if raw_transcript and session is not None:
                     await asyncio.shield(
                         _try_upload_transcript(
                             user_id,
@@ -1121,6 +1152,9 @@ async def stream_chat_completion_sdk(
         if sdk_cwd:
             _cleanup_sdk_tool_results(sdk_cwd)
 
+        # Release stream lock to allow new streams for this session
+        await lock.release()
+
 
 async def _try_upload_transcript(
     user_id: str,

autogpt_platform/backend/backend/copilot/service.py

@@ -352,7 +352,8 @@ async def assign_user_to_session(
     if not session:
         raise NotFoundError(f"Session {session_id} not found")
     session.user_id = user_id
-    return await upsert_chat_session(session)
+    session = await upsert_chat_session(session)
+    return session
 
 
 async def stream_chat_completion(
@@ -1563,7 +1564,11 @@ async def _yield_tool_call(
             await _mark_operation_completed(tool_call_id)
             # Mark stream registry task as failed if it was created
             try:
-                await stream_registry.mark_task_completed(task_id, status="failed")
+                await stream_registry.mark_task_completed(
+                    task_id,
+                    status="failed",
+                    error_message=f"Failed to setup tool {tool_name}: {e}",
+                )
             except Exception as mark_err:
                 logger.warning(f"Failed to mark task {task_id} as failed: {mark_err}")
             logger.error(
@@ -1731,7 +1736,11 @@ async def _execute_long_running_tool_with_streaming(
         session = await get_chat_session(session_id, user_id)
         if not session:
             logger.error(f"Session {session_id} not found for background tool")
-            await stream_registry.mark_task_completed(task_id, status="failed")
+            await stream_registry.mark_task_completed(
+                task_id,
+                status="failed",
+                error_message=f"Session {session_id} not found",
+            )
             return
 
         # Pass operation_id and task_id to the tool for async processing

autogpt_platform/backend/backend/copilot/stream_registry.py

@@ -644,6 +644,8 @@ async def _stream_listener(
 async def mark_task_completed(
     task_id: str,
     status: Literal["completed", "failed"] = "completed",
+    *,
+    error_message: str | None = None,
 ) -> bool:
     """Mark a task as completed and publish finish event.
 
@@ -654,6 +656,10 @@ async def mark_task_completed(
     Args:
         task_id: Task ID to mark as completed
         status: Final status ("completed" or "failed")
+        error_message: If provided and status="failed", publish a StreamError
+            before StreamFinish so connected clients see why the task ended.
+            If not provided, no StreamError is published (caller should publish
+            manually if needed to avoid duplicates).
 
     Returns:
         True if task was newly marked completed, False if already completed/failed
@@ -669,6 +675,17 @@ async def mark_task_completed(
         logger.debug(f"Task {task_id} already completed/failed, skipping")
         return False
 
+    # Publish error event before finish so connected clients know WHY the
+    # task ended. Only publish if caller provided an explicit error message
+    # to avoid duplicates with code paths that manually publish StreamError.
+    # This is best-effort — if it fails, the StreamFinish still ensures
+    # listeners clean up.
+    if status == "failed" and error_message:
+        try:
+            await publish_chunk(task_id, StreamError(errorText=error_message))
+        except Exception as e:
+            logger.warning(f"Failed to publish error event for task {task_id}: {e}")
+
     # THEN publish finish event (best-effort - listeners can detect via status polling)
     try:
         await publish_chunk(task_id, StreamFinish())
@@ -821,27 +838,6 @@ async def get_active_task_for_session(
                 if task_user_id and user_id != task_user_id:
                     continue
 
-                # Auto-expire stale tasks that exceeded stream_timeout
-                created_at_str = meta.get("created_at", "")
-                if created_at_str:
-                    try:
-                        created_at = datetime.fromisoformat(created_at_str)
-                        age_seconds = (
-                            datetime.now(timezone.utc) - created_at
-                        ).total_seconds()
-                        if age_seconds > config.stream_timeout:
-                            logger.warning(
-                                f"[TASK_LOOKUP] Auto-expiring stale task {task_id[:8]}... "
-                                f"(age={age_seconds:.0f}s > timeout={config.stream_timeout}s)"
-                            )
-                            await mark_task_completed(task_id, "failed")
-                            continue
-                    except (ValueError, TypeError) as exc:
-                        logger.warning(
-                            f"[TASK_LOOKUP] Failed to parse created_at "
-                            f"for task {task_id[:8]}...: {exc}"
-                        )
-
                 logger.info(
                     f"[TASK_LOOKUP] Found running task {task_id[:8]}... for session {session_id[:8]}..."
                 )

autogpt_platform/backend/backend/data/db_manager.py

@@ -303,7 +303,7 @@ class DatabaseManager(AppService):
     get_user_chat_sessions = _(chat_db.get_user_chat_sessions)
     get_user_session_count = _(chat_db.get_user_session_count)
     delete_chat_session = _(chat_db.delete_chat_session)
-    get_chat_session_message_count = _(chat_db.get_chat_session_message_count)
+    get_next_sequence = _(chat_db.get_next_sequence)
     update_tool_message_content = _(chat_db.update_tool_message_content)
 
 
@@ -473,5 +473,5 @@ class DatabaseManagerAsyncClient(AppServiceClient):
     get_user_chat_sessions = d.get_user_chat_sessions
     get_user_session_count = d.get_user_session_count
     delete_chat_session = d.delete_chat_session
-    get_chat_session_message_count = d.get_chat_session_message_count
+    get_next_sequence = d.get_next_sequence
     update_tool_message_content = d.update_tool_message_content

autogpt_platform/backend/backend/executor/cluster_lock.py

@@ -1,5 +1,6 @@
 """Redis-based distributed locking for cluster coordination."""
 
+import asyncio
 import logging
 import threading
 import time
@@ -7,6 +8,7 @@ from typing import TYPE_CHECKING
 
 if TYPE_CHECKING:
     from redis import Redis
+    from redis.asyncio import Redis as AsyncRedis
 
 logger = logging.getLogger(__name__)
 
@@ -126,3 +128,124 @@ class ClusterLock:
 
         with self._refresh_lock:
             self._last_refresh = 0.0
+
+
+class AsyncClusterLock:
+    """Async Redis-based distributed lock for preventing duplicate execution."""
+
+    def __init__(
+        self, redis: "AsyncRedis", key: str, owner_id: str, timeout: int = 300
+    ):
+        self.redis = redis
+        self.key = key
+        self.owner_id = owner_id
+        self.timeout = timeout
+        self._last_refresh = 0.0
+        self._refresh_lock = asyncio.Lock()
+
+    async def try_acquire(self) -> str | None:
+        """Try to acquire the lock.
+
+        Returns:
+            - owner_id (self.owner_id) if successfully acquired
+            - different owner_id if someone else holds the lock
+            - None if Redis is unavailable or other error
+        """
+        try:
+            success = await self.redis.set(
+                self.key, self.owner_id, nx=True, ex=self.timeout
+            )
+            if success:
+                async with self._refresh_lock:
+                    self._last_refresh = time.time()
+                return self.owner_id  # Successfully acquired
+
+            # Failed to acquire, get current owner
+            current_value = await self.redis.get(self.key)
+            if current_value:
+                current_owner = (
+                    current_value.decode("utf-8")
+                    if isinstance(current_value, bytes)
+                    else str(current_value)
+                )
+                return current_owner
+
+            # Key doesn't exist but we failed to set it - race condition or Redis issue
+            return None
+
+        except Exception as e:
+            logger.error(f"AsyncClusterLock.try_acquire failed for key {self.key}: {e}")
+            return None
+
+    async def refresh(self) -> bool:
+        """Refresh lock TTL if we still own it.
+
+        Rate limited to at most once every timeout/10 seconds (minimum 1 second).
+        During rate limiting, still verifies lock existence but skips TTL extension.
+        Setting _last_refresh to 0 bypasses rate limiting for testing.
+
+        Async-safe: uses asyncio.Lock to protect _last_refresh access.
+        """
+        # Calculate refresh interval: max(timeout // 10, 1)
+        refresh_interval = max(self.timeout // 10, 1)
+        current_time = time.time()
+
+        # Check if we're within the rate limit period (async-safe read)
+        # _last_refresh == 0 forces a refresh (bypasses rate limiting for testing)
+        async with self._refresh_lock:
+            last_refresh = self._last_refresh
+        is_rate_limited = (
+            last_refresh > 0 and (current_time - last_refresh) < refresh_interval
+        )
+
+        try:
+            # Always verify lock existence, even during rate limiting
+            current_value = await self.redis.get(self.key)
+            if not current_value:
+                async with self._refresh_lock:
+                    self._last_refresh = 0
+                return False
+
+            stored_owner = (
+                current_value.decode("utf-8")
+                if isinstance(current_value, bytes)
+                else str(current_value)
+            )
+            if stored_owner != self.owner_id:
+                async with self._refresh_lock:
+                    self._last_refresh = 0
+                return False
+
+            # If rate limited, return True but don't update TTL or timestamp
+            if is_rate_limited:
+                return True
+
+            # Perform actual refresh
+            if await self.redis.expire(self.key, self.timeout):
+                async with self._refresh_lock:
+                    self._last_refresh = current_time
+                return True
+
+            async with self._refresh_lock:
+                self._last_refresh = 0
+            return False
+
+        except Exception as e:
+            logger.error(f"AsyncClusterLock.refresh failed for key {self.key}: {e}")
+            async with self._refresh_lock:
+                self._last_refresh = 0
+            return False
+
+    async def release(self):
+        """Release the lock."""
+        async with self._refresh_lock:
+            if self._last_refresh == 0:
+                return
+
+        try:
+            await self.redis.delete(self.key)
+        except Exception:
+            pass
+
+        async with self._refresh_lock:
+            self._last_refresh = 0.0

autogpt_platform/backend/backend/util/settings.py

@@ -364,11 +364,11 @@ class Config(UpdateTrackingModel["Config"], BaseSettings):
     )
 
     agentgenerator_host: str = Field(
-        default="",
-        description="The host for the Agent Generator service (empty to use built-in)",
+        default="localhost",
+        description="The host for the Agent Generator service",
     )
     agentgenerator_port: int = Field(
-        default=8000,
+        default=8009,
         description="The port for the Agent Generator service",
     )
     agentgenerator_timeout: int = Field(