fix(backend/chat): Address PR review comments for Claude SDK integration

- Add StreamFinish after ErrorMessage in response adapter
- Fix str.replace to removeprefix in security hooks
- Apply max_context_messages limit as safety guard in history formatting
- Add empty prompt guard before sending to SDK
- Sanitize error messages to avoid exposing internal details
- Fix fire-and-forget asyncio.create_task by storing task reference
- Fix tool_calls population on assistant messages
- Rewrite Anthropic fallback to persist messages and merge consecutive roles
- Only use ANTHROPIC_API_KEY for fallback (not OpenRouter keys)
- Fix IndexError when tool result content list is empty

.gitignore

@@ -180,3 +180,4 @@ autogpt_platform/backend/settings.py
 .claude/settings.local.json
 CLAUDE.local.md
 /autogpt_platform/backend/logs
+.next

autogpt_platform/backend/.env.default

@@ -152,6 +152,7 @@ REPLICATE_API_KEY=
 REVID_API_KEY=
 SCREENSHOTONE_API_KEY=
 UNREAL_SPEECH_API_KEY=
+ELEVENLABS_API_KEY=
 
 # Data & Search Services
 E2B_API_KEY=

autogpt_platform/backend/.gitignore

@@ -19,3 +19,6 @@ load-tests/*.json
 load-tests/*.log
 load-tests/node_modules/*
 migrations/*/rollback*.sql
+
+# Workspace files
+workspaces/

autogpt_platform/backend/Dockerfile

@@ -62,10 +62,12 @@ ENV POETRY_HOME=/opt/poetry \
     DEBIAN_FRONTEND=noninteractive
 ENV PATH=/opt/poetry/bin:$PATH
 
-# Install Python without upgrading system-managed packages
+# Install Python, FFmpeg, and ImageMagick (required for video processing blocks)
 RUN apt-get update && apt-get install -y \
     python3.13 \
     python3-pip \
+    ffmpeg \
+    imagemagick \
     && rm -rf /var/lib/apt/lists/*
 
 # Copy only necessary files from builder

autogpt_platform/backend/backend/api/features/chat/completion_consumer.py

@@ -0,0 +1,368 @@
+"""Redis Streams consumer for operation completion messages.
+
+This module provides a consumer (ChatCompletionConsumer) that listens for
+completion notifications (OperationCompleteMessage) from external services
+(like Agent Generator) and triggers the appropriate stream registry and
+chat service updates via process_operation_success/process_operation_failure.
+
+Why Redis Streams instead of RabbitMQ?
+--------------------------------------
+While the project typically uses RabbitMQ for async task queues (e.g., execution
+queue), Redis Streams was chosen for chat completion notifications because:
+
+1. **Unified Infrastructure**: The SSE reconnection feature already uses Redis
+   Streams (via stream_registry) for message persistence and replay. Using Redis
+   Streams for completion notifications keeps all chat streaming infrastructure
+   in one system, simplifying operations and reducing cross-system coordination.
+
+2. **Message Replay**: Redis Streams support XREAD with arbitrary message IDs,
+   allowing consumers to replay missed messages after reconnection. This aligns
+   with the SSE reconnection pattern where clients can resume from last_message_id.
+
+3. **Consumer Groups with XAUTOCLAIM**: Redis consumer groups provide automatic
+   load balancing across pods with explicit message claiming (XAUTOCLAIM) for
+   recovering from dead consumers - ideal for the completion callback pattern.
+
+4. **Lower Latency**: For real-time SSE updates, Redis (already in-memory for
+   stream_registry) provides lower latency than an additional RabbitMQ hop.
+
+5. **Atomicity with Task State**: Completion processing often needs to update
+   task metadata stored in Redis. Keeping both in Redis enables simpler
+   transactional semantics without distributed coordination.
+
+The consumer uses Redis Streams with consumer groups for reliable message
+processing across multiple platform pods, with XAUTOCLAIM for reclaiming
+stale pending messages from dead consumers.
+"""
+
+import asyncio
+import logging
+import os
+import uuid
+from typing import Any
+
+import orjson
+from prisma import Prisma
+from pydantic import BaseModel
+from redis.exceptions import ResponseError
+
+from backend.data.redis_client import get_redis_async
+
+from . import stream_registry
+from .completion_handler import process_operation_failure, process_operation_success
+from .config import ChatConfig
+
+logger = logging.getLogger(__name__)
+config = ChatConfig()
+
+
+class OperationCompleteMessage(BaseModel):
+    """Message format for operation completion notifications."""
+
+    operation_id: str
+    task_id: str
+    success: bool
+    result: dict | str | None = None
+    error: str | None = None
+
+
+class ChatCompletionConsumer:
+    """Consumer for chat operation completion messages from Redis Streams.
+
+    This consumer initializes its own Prisma client in start() to ensure
+    database operations work correctly within this async context.
+
+    Uses Redis consumer groups to allow multiple platform pods to consume
+    messages reliably with automatic redelivery on failure.
+    """
+
+    def __init__(self):
+        self._consumer_task: asyncio.Task | None = None
+        self._running = False
+        self._prisma: Prisma | None = None
+        self._consumer_name = f"consumer-{uuid.uuid4().hex[:8]}"
+
+    async def start(self) -> None:
+        """Start the completion consumer."""
+        if self._running:
+            logger.warning("Completion consumer already running")
+            return
+
+        # Create consumer group if it doesn't exist
+        try:
+            redis = await get_redis_async()
+            await redis.xgroup_create(
+                config.stream_completion_name,
+                config.stream_consumer_group,
+                id="0",
+                mkstream=True,
+            )
+            logger.info(
+                f"Created consumer group '{config.stream_consumer_group}' "
+                f"on stream '{config.stream_completion_name}'"
+            )
+        except ResponseError as e:
+            if "BUSYGROUP" in str(e):
+                logger.debug(
+                    f"Consumer group '{config.stream_consumer_group}' already exists"
+                )
+            else:
+                raise
+
+        self._running = True
+        self._consumer_task = asyncio.create_task(self._consume_messages())
+        logger.info(
+            f"Chat completion consumer started (consumer: {self._consumer_name})"
+        )
+
+    async def _ensure_prisma(self) -> Prisma:
+        """Lazily initialize Prisma client on first use."""
+        if self._prisma is None:
+            database_url = os.getenv("DATABASE_URL", "postgresql://localhost:5432")
+            self._prisma = Prisma(datasource={"url": database_url})
+            await self._prisma.connect()
+            logger.info("[COMPLETION] Consumer Prisma client connected (lazy init)")
+        return self._prisma
+
+    async def stop(self) -> None:
+        """Stop the completion consumer."""
+        self._running = False
+
+        if self._consumer_task:
+            self._consumer_task.cancel()
+            try:
+                await self._consumer_task
+            except asyncio.CancelledError:
+                pass
+            self._consumer_task = None
+
+        if self._prisma:
+            await self._prisma.disconnect()
+            self._prisma = None
+            logger.info("[COMPLETION] Consumer Prisma client disconnected")
+
+        logger.info("Chat completion consumer stopped")
+
+    async def _consume_messages(self) -> None:
+        """Main message consumption loop with retry logic."""
+        max_retries = 10
+        retry_delay = 5  # seconds
+        retry_count = 0
+        block_timeout = 5000  # milliseconds
+
+        while self._running and retry_count < max_retries:
+            try:
+                redis = await get_redis_async()
+
+                # Reset retry count on successful connection
+                retry_count = 0
+
+                while self._running:
+                    # First, claim any stale pending messages from dead consumers
+                    # Redis does NOT auto-redeliver pending messages; we must explicitly
+                    # claim them using XAUTOCLAIM
+                    try:
+                        claimed_result = await redis.xautoclaim(
+                            name=config.stream_completion_name,
+                            groupname=config.stream_consumer_group,
+                            consumername=self._consumer_name,
+                            min_idle_time=config.stream_claim_min_idle_ms,
+                            start_id="0-0",
+                            count=10,
+                        )
+                        # xautoclaim returns: (next_start_id, [(id, data), ...], [deleted_ids])
+                        if claimed_result and len(claimed_result) >= 2:
+                            claimed_entries = claimed_result[1]
+                            if claimed_entries:
+                                logger.info(
+                                    f"Claimed {len(claimed_entries)} stale pending messages"
+                                )
+                                for entry_id, data in claimed_entries:
+                                    if not self._running:
+                                        return
+                                    await self._process_entry(redis, entry_id, data)
+                    except Exception as e:
+                        logger.warning(f"XAUTOCLAIM failed (non-fatal): {e}")
+
+                    # Read new messages from the stream
+                    messages = await redis.xreadgroup(
+                        groupname=config.stream_consumer_group,
+                        consumername=self._consumer_name,
+                        streams={config.stream_completion_name: ">"},
+                        block=block_timeout,
+                        count=10,
+                    )
+
+                    if not messages:
+                        continue
+
+                    for stream_name, entries in messages:
+                        for entry_id, data in entries:
+                            if not self._running:
+                                return
+                            await self._process_entry(redis, entry_id, data)
+
+            except asyncio.CancelledError:
+                logger.info("Consumer cancelled")
+                return
+            except Exception as e:
+                retry_count += 1
+                logger.error(
+                    f"Consumer error (retry {retry_count}/{max_retries}): {e}",
+                    exc_info=True,
+                )
+                if self._running and retry_count < max_retries:
+                    await asyncio.sleep(retry_delay)
+                else:
+                    logger.error("Max retries reached, stopping consumer")
+                    return
+
+    async def _process_entry(
+        self, redis: Any, entry_id: str, data: dict[str, Any]
+    ) -> None:
+        """Process a single stream entry and acknowledge it on success.
+
+        Args:
+            redis: Redis client connection
+            entry_id: The stream entry ID
+            data: The entry data dict
+        """
+        try:
+            # Handle the message
+            message_data = data.get("data")
+            if message_data:
+                await self._handle_message(
+                    message_data.encode()
+                    if isinstance(message_data, str)
+                    else message_data
+                )
+
+            # Acknowledge the message after successful processing
+            await redis.xack(
+                config.stream_completion_name,
+                config.stream_consumer_group,
+                entry_id,
+            )
+        except Exception as e:
+            logger.error(
+                f"Error processing completion message {entry_id}: {e}",
+                exc_info=True,
+            )
+            # Message remains in pending state and will be claimed by
+            # XAUTOCLAIM after min_idle_time expires
+
+    async def _handle_message(self, body: bytes) -> None:
+        """Handle a completion message using our own Prisma client."""
+        try:
+            data = orjson.loads(body)
+            message = OperationCompleteMessage(**data)
+        except Exception as e:
+            logger.error(f"Failed to parse completion message: {e}")
+            return
+
+        logger.info(
+            f"[COMPLETION] Received completion for operation {message.operation_id} "
+            f"(task_id={message.task_id}, success={message.success})"
+        )
+
+        # Find task in registry
+        task = await stream_registry.find_task_by_operation_id(message.operation_id)
+        if task is None:
+            task = await stream_registry.get_task(message.task_id)
+
+        if task is None:
+            logger.warning(
+                f"[COMPLETION] Task not found for operation {message.operation_id} "
+                f"(task_id={message.task_id})"
+            )
+            return
+
+        logger.info(
+            f"[COMPLETION] Found task: task_id={task.task_id}, "
+            f"session_id={task.session_id}, tool_call_id={task.tool_call_id}"
+        )
+
+        # Guard against empty task fields
+        if not task.task_id or not task.session_id or not task.tool_call_id:
+            logger.error(
+                f"[COMPLETION] Task has empty critical fields! "
+                f"task_id={task.task_id!r}, session_id={task.session_id!r}, "
+                f"tool_call_id={task.tool_call_id!r}"
+            )
+            return
+
+        if message.success:
+            await self._handle_success(task, message)
+        else:
+            await self._handle_failure(task, message)
+
+    async def _handle_success(
+        self,
+        task: stream_registry.ActiveTask,
+        message: OperationCompleteMessage,
+    ) -> None:
+        """Handle successful operation completion."""
+        prisma = await self._ensure_prisma()
+        await process_operation_success(task, message.result, prisma)
+
+    async def _handle_failure(
+        self,
+        task: stream_registry.ActiveTask,
+        message: OperationCompleteMessage,
+    ) -> None:
+        """Handle failed operation completion."""
+        prisma = await self._ensure_prisma()
+        await process_operation_failure(task, message.error, prisma)
+
+
+# Module-level consumer instance
+_consumer: ChatCompletionConsumer | None = None
+
+
+async def start_completion_consumer() -> None:
+    """Start the global completion consumer."""
+    global _consumer
+    if _consumer is None:
+        _consumer = ChatCompletionConsumer()
+    await _consumer.start()
+
+
+async def stop_completion_consumer() -> None:
+    """Stop the global completion consumer."""
+    global _consumer
+    if _consumer:
+        await _consumer.stop()
+        _consumer = None
+
+
+async def publish_operation_complete(
+    operation_id: str,
+    task_id: str,
+    success: bool,
+    result: dict | str | None = None,
+    error: str | None = None,
+) -> None:
+    """Publish an operation completion message to Redis Streams.
+
+    Args:
+        operation_id: The operation ID that completed.
+        task_id: The task ID associated with the operation.
+        success: Whether the operation succeeded.
+        result: The result data (for success).
+        error: The error message (for failure).
+    """
+    message = OperationCompleteMessage(
+        operation_id=operation_id,
+        task_id=task_id,
+        success=success,
+        result=result,
+        error=error,
+    )
+
+    redis = await get_redis_async()
+    await redis.xadd(
+        config.stream_completion_name,
+        {"data": message.model_dump_json()},
+        maxlen=config.stream_max_length,
+    )
+    logger.info(f"Published completion for operation {operation_id}")

autogpt_platform/backend/backend/api/features/chat/completion_handler.py

@@ -0,0 +1,344 @@
+"""Shared completion handling for operation success and failure.
+
+This module provides common logic for handling operation completion from both:
+- The Redis Streams consumer (completion_consumer.py)
+- The HTTP webhook endpoint (routes.py)
+"""
+
+import logging
+from typing import Any
+
+import orjson
+from prisma import Prisma
+
+from . import service as chat_service
+from . import stream_registry
+from .response_model import StreamError, StreamToolOutputAvailable
+from .tools.models import ErrorResponse
+
+logger = logging.getLogger(__name__)
+
+# Tools that produce agent_json that needs to be saved to library
+AGENT_GENERATION_TOOLS = {"create_agent", "edit_agent"}
+
+# Keys that should be stripped from agent_json when returning in error responses
+SENSITIVE_KEYS = frozenset(
+    {
+        "api_key",
+        "apikey",
+        "api_secret",
+        "password",
+        "secret",
+        "credentials",
+        "credential",
+        "token",
+        "access_token",
+        "refresh_token",
+        "private_key",
+        "privatekey",
+        "auth",
+        "authorization",
+    }
+)
+
+
+def _sanitize_agent_json(obj: Any) -> Any:
+    """Recursively sanitize agent_json by removing sensitive keys.
+
+    Args:
+        obj: The object to sanitize (dict, list, or primitive)
+
+    Returns:
+        Sanitized copy with sensitive keys removed/redacted
+    """
+    if isinstance(obj, dict):
+        return {
+            k: "[REDACTED]" if k.lower() in SENSITIVE_KEYS else _sanitize_agent_json(v)
+            for k, v in obj.items()
+        }
+    elif isinstance(obj, list):
+        return [_sanitize_agent_json(item) for item in obj]
+    else:
+        return obj
+
+
+class ToolMessageUpdateError(Exception):
+    """Raised when updating a tool message in the database fails."""
+
+    pass
+
+
+async def _update_tool_message(
+    session_id: str,
+    tool_call_id: str,
+    content: str,
+    prisma_client: Prisma | None,
+) -> None:
+    """Update tool message in database.
+
+    Args:
+        session_id: The session ID
+        tool_call_id: The tool call ID to update
+        content: The new content for the message
+        prisma_client: Optional Prisma client. If None, uses chat_service.
+
+    Raises:
+        ToolMessageUpdateError: If the database update fails. The caller should
+            handle this to avoid marking the task as completed with inconsistent state.
+    """
+    try:
+        if prisma_client:
+            # Use provided Prisma client (for consumer with its own connection)
+            updated_count = await prisma_client.chatmessage.update_many(
+                where={
+                    "sessionId": session_id,
+                    "toolCallId": tool_call_id,
+                },
+                data={"content": content},
+            )
+            # Check if any rows were updated - 0 means message not found
+            if updated_count == 0:
+                raise ToolMessageUpdateError(
+                    f"No message found with tool_call_id={tool_call_id} in session {session_id}"
+                )
+        else:
+            # Use service function (for webhook endpoint)
+            await chat_service._update_pending_operation(
+                session_id=session_id,
+                tool_call_id=tool_call_id,
+                result=content,
+            )
+    except ToolMessageUpdateError:
+        raise
+    except Exception as e:
+        logger.error(f"[COMPLETION] Failed to update tool message: {e}", exc_info=True)
+        raise ToolMessageUpdateError(
+            f"Failed to update tool message for tool_call_id={tool_call_id}: {e}"
+        ) from e
+
+
+def serialize_result(result: dict | list | str | int | float | bool | None) -> str:
+    """Serialize result to JSON string with sensible defaults.
+
+    Args:
+        result: The result to serialize. Can be a dict, list, string,
+            number, boolean, or None.
+
+    Returns:
+        JSON string representation of the result. Returns '{"status": "completed"}'
+        only when result is explicitly None.
+    """
+    if isinstance(result, str):
+        return result
+    if result is None:
+        return '{"status": "completed"}'
+    return orjson.dumps(result).decode("utf-8")
+
+
+async def _save_agent_from_result(
+    result: dict[str, Any],
+    user_id: str | None,
+    tool_name: str,
+) -> dict[str, Any]:
+    """Save agent to library if result contains agent_json.
+
+    Args:
+        result: The result dict that may contain agent_json
+        user_id: The user ID to save the agent for
+        tool_name: The tool name (create_agent or edit_agent)
+
+    Returns:
+        Updated result dict with saved agent details, or original result if no agent_json
+    """
+    if not user_id:
+        logger.warning("[COMPLETION] Cannot save agent: no user_id in task")
+        return result
+
+    agent_json = result.get("agent_json")
+    if not agent_json:
+        logger.warning(
+            f"[COMPLETION] {tool_name} completed but no agent_json in result"
+        )
+        return result
+
+    try:
+        from .tools.agent_generator import save_agent_to_library
+
+        is_update = tool_name == "edit_agent"
+        created_graph, library_agent = await save_agent_to_library(
+            agent_json, user_id, is_update=is_update
+        )
+
+        logger.info(
+            f"[COMPLETION] Saved agent '{created_graph.name}' to library "
+            f"(graph_id={created_graph.id}, library_agent_id={library_agent.id})"
+        )
+
+        # Return a response similar to AgentSavedResponse
+        return {
+            "type": "agent_saved",
+            "message": f"Agent '{created_graph.name}' has been saved to your library!",
+            "agent_id": created_graph.id,
+            "agent_name": created_graph.name,
+            "library_agent_id": library_agent.id,
+            "library_agent_link": f"/library/agents/{library_agent.id}",
+            "agent_page_link": f"/build?flowID={created_graph.id}",
+        }
+    except Exception as e:
+        logger.error(
+            f"[COMPLETION] Failed to save agent to library: {e}",
+            exc_info=True,
+        )
+        # Return error but don't fail the whole operation
+        # Sanitize agent_json to remove sensitive keys before returning
+        return {
+            "type": "error",
+            "message": f"Agent was generated but failed to save: {str(e)}",
+            "error": str(e),
+            "agent_json": _sanitize_agent_json(agent_json),
+        }
+
+
+async def process_operation_success(
+    task: stream_registry.ActiveTask,
+    result: dict | str | None,
+    prisma_client: Prisma | None = None,
+) -> None:
+    """Handle successful operation completion.
+
+    Publishes the result to the stream registry, updates the database,
+    generates LLM continuation, and marks the task as completed.
+
+    Args:
+        task: The active task that completed
+        result: The result data from the operation
+        prisma_client: Optional Prisma client for database operations.
+            If None, uses chat_service._update_pending_operation instead.
+
+    Raises:
+        ToolMessageUpdateError: If the database update fails. The task will be
+            marked as failed instead of completed to avoid inconsistent state.
+    """
+    # For agent generation tools, save the agent to library
+    if task.tool_name in AGENT_GENERATION_TOOLS and isinstance(result, dict):
+        result = await _save_agent_from_result(result, task.user_id, task.tool_name)
+
+    # Serialize result for output (only substitute default when result is exactly None)
+    result_output = result if result is not None else {"status": "completed"}
+    output_str = (
+        result_output
+        if isinstance(result_output, str)
+        else orjson.dumps(result_output).decode("utf-8")
+    )
+
+    # Publish result to stream registry
+    await stream_registry.publish_chunk(
+        task.task_id,
+        StreamToolOutputAvailable(
+            toolCallId=task.tool_call_id,
+            toolName=task.tool_name,
+            output=output_str,
+            success=True,
+        ),
+    )
+
+    # Update pending operation in database
+    # If this fails, we must not continue to mark the task as completed
+    result_str = serialize_result(result)
+    try:
+        await _update_tool_message(
+            session_id=task.session_id,
+            tool_call_id=task.tool_call_id,
+            content=result_str,
+            prisma_client=prisma_client,
+        )
+    except ToolMessageUpdateError:
+        # DB update failed - mark task as failed to avoid inconsistent state
+        logger.error(
+            f"[COMPLETION] DB update failed for task {task.task_id}, "
+            "marking as failed instead of completed"
+        )
+        await stream_registry.publish_chunk(
+            task.task_id,
+            StreamError(errorText="Failed to save operation result to database"),
+        )
+        await stream_registry.mark_task_completed(task.task_id, status="failed")
+        raise
+
+    # Generate LLM continuation with streaming
+    try:
+        await chat_service._generate_llm_continuation_with_streaming(
+            session_id=task.session_id,
+            user_id=task.user_id,
+            task_id=task.task_id,
+        )
+    except Exception as e:
+        logger.error(
+            f"[COMPLETION] Failed to generate LLM continuation: {e}",
+            exc_info=True,
+        )
+
+    # Mark task as completed and release Redis lock
+    await stream_registry.mark_task_completed(task.task_id, status="completed")
+    try:
+        await chat_service._mark_operation_completed(task.tool_call_id)
+    except Exception as e:
+        logger.error(f"[COMPLETION] Failed to mark operation completed: {e}")
+
+    logger.info(
+        f"[COMPLETION] Successfully processed completion for task {task.task_id}"
+    )
+
+
+async def process_operation_failure(
+    task: stream_registry.ActiveTask,
+    error: str | None,
+    prisma_client: Prisma | None = None,
+) -> None:
+    """Handle failed operation completion.
+
+    Publishes the error to the stream registry, updates the database with
+    the error response, and marks the task as failed.
+
+    Args:
+        task: The active task that failed
+        error: The error message from the operation
+        prisma_client: Optional Prisma client for database operations.
+            If None, uses chat_service._update_pending_operation instead.
+    """
+    error_msg = error or "Operation failed"
+
+    # Publish error to stream registry
+    await stream_registry.publish_chunk(
+        task.task_id,
+        StreamError(errorText=error_msg),
+    )
+
+    # Update pending operation with error
+    # If this fails, we still continue to mark the task as failed
+    error_response = ErrorResponse(
+        message=error_msg,
+        error=error,
+    )
+    try:
+        await _update_tool_message(
+            session_id=task.session_id,
+            tool_call_id=task.tool_call_id,
+            content=error_response.model_dump_json(),
+            prisma_client=prisma_client,
+        )
+    except ToolMessageUpdateError:
+        # DB update failed - log but continue with cleanup
+        logger.error(
+            f"[COMPLETION] DB update failed while processing failure for task {task.task_id}, "
+            "continuing with cleanup"
+        )
+
+    # Mark task as failed and release Redis lock
+    await stream_registry.mark_task_completed(task.task_id, status="failed")
+    try:
+        await chat_service._mark_operation_completed(task.tool_call_id)
+    except Exception as e:
+        logger.error(f"[COMPLETION] Failed to mark operation completed: {e}")
+
+    logger.info(f"[COMPLETION] Processed failure for task {task.task_id}: {error_msg}")

autogpt_platform/backend/backend/api/features/chat/config.py

@@ -11,7 +11,7 @@ class ChatConfig(BaseSettings):
 
     # OpenAI API Configuration
     model: str = Field(
-        default="anthropic/claude-opus-4.5", description="Default model to use"
+        default="anthropic/claude-opus-4.6", description="Default model to use"
     )
     title_model: str = Field(
         default="openai/gpt-4o-mini",
@@ -27,12 +27,20 @@ class ChatConfig(BaseSettings):
     session_ttl: int = Field(default=43200, description="Session TTL in seconds")
 
     # Streaming Configuration
+    # Note: When using Claude Agent SDK, context management is handled automatically
+    # via the SDK's built-in compaction. This is mainly used for the fallback path.
     max_context_messages: int = Field(
-        default=50, ge=1, le=200, description="Maximum context messages"
+        default=100,
+        ge=1,
+        le=500,
+        description="Max context messages (SDK handles compaction automatically)",
     )
 
     stream_timeout: int = Field(default=300, description="Stream timeout in seconds")
-    max_retries: int = Field(default=3, description="Maximum number of retries")
+    max_retries: int = Field(
+        default=3,
+        description="Max retries for fallback path (SDK handles retries internally)",
+    )
     max_agent_runs: int = Field(default=30, description="Maximum number of agent runs")
     max_agent_schedules: int = Field(
         default=30, description="Maximum number of agent schedules"
@@ -44,6 +52,48 @@ class ChatConfig(BaseSettings):
         description="TTL in seconds for long-running operation tracking in Redis (safety net if pod dies)",
     )
 
+    # Stream registry configuration for SSE reconnection
+    stream_ttl: int = Field(
+        default=3600,
+        description="TTL in seconds for stream data in Redis (1 hour)",
+    )
+    stream_max_length: int = Field(
+        default=10000,
+        description="Maximum number of messages to store per stream",
+    )
+
+    # Redis Streams configuration for completion consumer
+    stream_completion_name: str = Field(
+        default="chat:completions",
+        description="Redis Stream name for operation completions",
+    )
+    stream_consumer_group: str = Field(
+        default="chat_consumers",
+        description="Consumer group name for completion stream",
+    )
+    stream_claim_min_idle_ms: int = Field(
+        default=60000,
+        description="Minimum idle time in milliseconds before claiming pending messages from dead consumers",
+    )
+
+    # Redis key prefixes for stream registry
+    task_meta_prefix: str = Field(
+        default="chat:task:meta:",
+        description="Prefix for task metadata hash keys",
+    )
+    task_stream_prefix: str = Field(
+        default="chat:stream:",
+        description="Prefix for task message stream keys",
+    )
+    task_op_prefix: str = Field(
+        default="chat:task:op:",
+        description="Prefix for operation ID to task ID mapping keys",
+    )
+    internal_api_key: str | None = Field(
+        default=None,
+        description="API key for internal webhook callbacks (env: CHAT_INTERNAL_API_KEY)",
+    )
+
     # Langfuse Prompt Management Configuration
     # Note: Langfuse credentials are in Settings().secrets (settings.py)
     langfuse_prompt_name: str = Field(
@@ -51,6 +101,12 @@ class ChatConfig(BaseSettings):
         description="Name of the prompt in Langfuse to fetch",
     )
 
+    # Claude Agent SDK Configuration
+    use_claude_agent_sdk: bool = Field(
+        default=True,
+        description="Use Claude Agent SDK for chat completions",
+    )
+
     @field_validator("api_key", mode="before")
     @classmethod
     def get_api_key(cls, v):
@@ -82,6 +138,25 @@ class ChatConfig(BaseSettings):
                 v = "https://openrouter.ai/api/v1"
         return v
 
+    @field_validator("internal_api_key", mode="before")
+    @classmethod
+    def get_internal_api_key(cls, v):
+        """Get internal API key from environment if not provided."""
+        if v is None:
+            v = os.getenv("CHAT_INTERNAL_API_KEY")
+        return v
+
+    @field_validator("use_claude_agent_sdk", mode="before")
+    @classmethod
+    def get_use_claude_agent_sdk(cls, v):
+        """Get use_claude_agent_sdk from environment if not provided."""
+        # Check environment variable - default to True if not set
+        env_val = os.getenv("CHAT_USE_CLAUDE_AGENT_SDK", "").lower()
+        if env_val:
+            return env_val in ("true", "1", "yes", "on")
+        # Default to True (SDK enabled by default)
+        return True if v is None else v
+
     # Prompt paths for different contexts
     PROMPT_PATHS: dict[str, str] = {
         "default": "prompts/chat_system.md",

autogpt_platform/backend/backend/api/features/chat/model.py

@@ -273,9 +273,8 @@ async def _get_session_from_cache(session_id: str) -> ChatSession | None:
     try:
         session = ChatSession.model_validate_json(raw_session)
         logger.info(
-            f"Loading session {session_id} from cache: "
+            f"[CACHE] Loaded session {session_id}: {len(session.messages)} messages, "
-            f"message_count={len(session.messages)}, "
+            f"last_roles={[m.role for m in session.messages[-3:]]}"  # Last 3 roles
-            f"roles={[m.role for m in session.messages]}"
         )
         return session
     except Exception as e:
@@ -317,11 +316,9 @@ async def _get_session_from_db(session_id: str) -> ChatSession | None:
         return None
 
     messages = prisma_session.Messages
-    logger.info(
+    logger.debug(
-        f"Loading session {session_id} from DB: "
+        f"[DB] Loaded session {session_id}: {len(messages) if messages else 0} messages, "
-        f"has_messages={messages is not None}, "
+        f"roles={[m.role for m in messages[-3:]] if messages else []}"  # Last 3 roles
-        f"message_count={len(messages) if messages else 0}, "
-        f"roles={[m.role for m in messages] if messages else []}"
     )
 
     return ChatSession.from_db(prisma_session, messages)
@@ -372,10 +369,9 @@ async def _save_session_to_db(
                     "function_call": msg.function_call,
                 }
             )
-        logger.info(
+        logger.debug(
-            f"Saving {len(new_messages)} new messages to DB for session {session.session_id}: "
+            f"[DB] Saving {len(new_messages)} messages to session {session.session_id}, "
-            f"roles={[m['role'] for m in messages_data]}, "
+            f"roles={[m['role'] for m in messages_data]}"
-            f"start_sequence={existing_message_count}"
         )
         await chat_db.add_chat_messages_batch(
             session_id=session.session_id,
@@ -415,7 +411,7 @@ async def get_chat_session(
         logger.warning(f"Unexpected cache error for session {session_id}: {e}")
 
     # Fall back to database
-    logger.info(f"Session {session_id} not in cache, checking database")
+    logger.debug(f"Session {session_id} not in cache, checking database")
     session = await _get_session_from_db(session_id)
 
     if session is None:
@@ -432,7 +428,6 @@ async def get_chat_session(
     # Cache the session from DB
     try:
         await _cache_session(session)
-        logger.info(f"Cached session {session_id} from database")
     except Exception as e:
         logger.warning(f"Failed to cache session {session_id}: {e}")
 
@@ -603,13 +598,19 @@ async def update_session_title(session_id: str, title: str) -> bool:
             logger.warning(f"Session {session_id} not found for title update")
             return False
 
-        # Invalidate cache so next fetch gets updated title
+        # Update title in cache if it exists (instead of invalidating).
+        # This prevents race conditions where cache invalidation causes
+        # the frontend to see stale DB data while streaming is still in progress.
         try:
-            redis_key = _get_session_cache_key(session_id)
+            cached = await _get_session_from_cache(session_id)
-            async_redis = await get_redis_async()
+            if cached:
-            await async_redis.delete(redis_key)
+                cached.title = title
+                await _cache_session(cached)
         except Exception as e:
-            logger.warning(f"Failed to invalidate cache for session {session_id}: {e}")
+            # Not critical - title will be correct on next full cache refresh
+            logger.warning(
+                f"Failed to update title in cache for session {session_id}: {e}"
+            )
 
         return True
     except Exception as e:

autogpt_platform/backend/backend/api/features/chat/response_model.py

@@ -52,6 +52,10 @@ class StreamStart(StreamBaseResponse):
 
     type: ResponseType = ResponseType.START
     messageId: str = Field(..., description="Unique message ID")
+    taskId: str | None = Field(
+        default=None,
+        description="Task ID for SSE reconnection. Clients can reconnect using GET /tasks/{taskId}/stream",
+    )
 
 
 class StreamFinish(StreamBaseResponse):

autogpt_platform/backend/backend/api/features/chat/routes.py

@@ -1,19 +1,33 @@
 """Chat API routes for chat session management and streaming via SSE."""
 
+import asyncio
 import logging
+import uuid as uuid_module
 from collections.abc import AsyncGenerator
 from typing import Annotated
 
 from autogpt_libs import auth
-from fastapi import APIRouter, Depends, Query, Security
+from fastapi import APIRouter, Depends, Header, HTTPException, Query, Security
 from fastapi.responses import StreamingResponse
 from pydantic import BaseModel
 
 from backend.util.exceptions import NotFoundError
 
 from . import service as chat_service
+from . import stream_registry
+from .completion_handler import process_operation_failure, process_operation_success
 from .config import ChatConfig
-from .model import ChatSession, create_chat_session, get_chat_session, get_user_sessions
+from .model import (
+    ChatMessage,
+    ChatSession,
+    create_chat_session,
+    get_chat_session,
+    get_user_sessions,
+    upsert_chat_session,
+)
+from .response_model import StreamFinish, StreamHeartbeat, StreamStart
+from .sdk import service as sdk_service
+from .tracking import track_user_message
 
 config = ChatConfig()
 
@@ -55,6 +69,15 @@ class CreateSessionResponse(BaseModel):
     user_id: str | None
 
 
+class ActiveStreamInfo(BaseModel):
+    """Information about an active stream for reconnection."""
+
+    task_id: str
+    last_message_id: str  # Redis Stream message ID for resumption
+    operation_id: str  # Operation ID for completion tracking
+    tool_name: str  # Name of the tool being executed
+
+
 class SessionDetailResponse(BaseModel):
     """Response model providing complete details for a chat session, including messages."""
 
@@ -63,6 +86,7 @@ class SessionDetailResponse(BaseModel):
     updated_at: str
     user_id: str | None
     messages: list[dict]
+    active_stream: ActiveStreamInfo | None = None  # Present if stream is still active
 
 
 class SessionSummaryResponse(BaseModel):
@@ -81,6 +105,14 @@ class ListSessionsResponse(BaseModel):
     total: int
 
 
+class OperationCompleteRequest(BaseModel):
+    """Request model for external completion webhook."""
+
+    success: bool
+    result: dict | str | None = None
+    error: str | None = None
+
+
 # ========== Routes ==========
 
 
@@ -166,13 +198,14 @@ async def get_session(
     Retrieve the details of a specific chat session.
 
     Looks up a chat session by ID for the given user (if authenticated) and returns all session data including messages.
+    If there's an active stream for this session, returns the task_id for reconnection.
 
     Args:
         session_id: The unique identifier for the desired chat session.
         user_id: The optional authenticated user ID, or None for anonymous access.
 
     Returns:
-        SessionDetailResponse: Details for the requested session, or None if not found.
+        SessionDetailResponse: Details for the requested session, including active_stream info if applicable.
 
     """
     session = await get_chat_session(session_id, user_id)
@@ -180,11 +213,32 @@ async def get_session(
         raise NotFoundError(f"Session {session_id} not found.")
 
     messages = [message.model_dump() for message in session.messages]
-    logger.info(
+
-        f"Returning session {session_id}: "
+    # Check if there's an active stream for this session
-        f"message_count={len(messages)}, "
+    active_stream_info = None
-        f"roles={[m.get('role') for m in messages]}"
+    active_task, last_message_id = await stream_registry.get_active_task_for_session(
+        session_id, user_id
     )
+    logger.info(
+        f"[GET_SESSION] session={session_id}, active_task={active_task is not None}, "
+        f"msg_count={len(messages)}, last_role={messages[-1].get('role') if messages else 'none'}"
+    )
+    if active_task:
+        # Filter out the in-progress assistant message from the session response.
+        # The client will receive the complete assistant response through the SSE
+        # stream replay instead, preventing duplicate content.
+        if messages and messages[-1].get("role") == "assistant":
+            messages = messages[:-1]
+
+        # Use "0-0" as last_message_id to replay the stream from the beginning.
+        # Since we filtered out the cached assistant message, the client needs
+        # the full stream to reconstruct the response.
+        active_stream_info = ActiveStreamInfo(
+            task_id=active_task.task_id,
+            last_message_id="0-0",
+            operation_id=active_task.operation_id,
+            tool_name=active_task.tool_name,
+        )
 
     return SessionDetailResponse(
         id=session.session_id,
@@ -192,6 +246,7 @@ async def get_session(
         updated_at=session.updated_at.isoformat(),
         user_id=session.user_id or None,
         messages=messages,
+        active_stream=active_stream_info,
     )
 
 
@@ -211,49 +266,147 @@ async def stream_chat_post(
       - Tool call UI elements (if invoked)
       - Tool execution results
 
+    The AI generation runs in a background task that continues even if the client disconnects.
+    All chunks are written to Redis for reconnection support. If the client disconnects,
+    they can reconnect using GET /tasks/{task_id}/stream to resume from where they left off.
+
     Args:
         session_id: The chat session identifier to associate with the streamed messages.
         request: Request body containing message, is_user_message, and optional context.
         user_id: Optional authenticated user ID.
     Returns:
-        StreamingResponse: SSE-formatted response chunks.
+        StreamingResponse: SSE-formatted response chunks. First chunk is a "start" event
+        containing the task_id for reconnection.
 
     """
     session = await _validate_and_get_session(session_id, user_id)
 
-    async def event_generator() -> AsyncGenerator[str, None]:
+    # Add user message to session BEFORE creating task to avoid race condition
-        chunk_count = 0
+    # where GET_SESSION sees the task as "running" but the message isn't saved yet
-        first_chunk_type: str | None = None
+    if request.message:
-        async for chunk in chat_service.stream_chat_completion(
+        session.messages.append(
-            session_id,
+            ChatMessage(
-            request.message,
+                role="user" if request.is_user_message else "assistant",
-            is_user_message=request.is_user_message,
+                content=request.message,
-            user_id=user_id,
+            )
-            session=session,  # Pass pre-fetched session to avoid double-fetch
-            context=request.context,
-        ):
-            if chunk_count < 3:
-                logger.info(
-                    "Chat stream chunk",
-                    extra={
-                        "session_id": session_id,
-                        "chunk_type": str(chunk.type),
-                    },
-                )
-            if not first_chunk_type:
-                first_chunk_type = str(chunk.type)
-            chunk_count += 1
-            yield chunk.to_sse()
-        logger.info(
-            "Chat stream completed",
-            extra={
-                "session_id": session_id,
-                "chunk_count": chunk_count,
-                "first_chunk_type": first_chunk_type,
-            },
         )
-        # AI SDK protocol termination
+        if request.is_user_message:
-        yield "data: [DONE]\n\n"
+            track_user_message(
+                user_id=user_id,
+                session_id=session_id,
+                message_length=len(request.message),
+            )
+        logger.info(
+            f"[STREAM] Saving user message to session {session_id}, "
+            f"msg_count={len(session.messages)}"
+        )
+        session = await upsert_chat_session(session)
+        logger.info(f"[STREAM] User message saved for session {session_id}")
+
+    # Create a task in the stream registry for reconnection support
+    task_id = str(uuid_module.uuid4())
+    operation_id = str(uuid_module.uuid4())
+    await stream_registry.create_task(
+        task_id=task_id,
+        session_id=session_id,
+        user_id=user_id,
+        tool_call_id="chat_stream",  # Not a tool call, but needed for the model
+        tool_name="chat",
+        operation_id=operation_id,
+    )
+
+    # Background task that runs the AI generation independently of SSE connection
+    async def run_ai_generation():
+        chunk_count = 0
+        try:
+            # Emit a start event with task_id for reconnection
+            start_chunk = StreamStart(messageId=task_id, taskId=task_id)
+            await stream_registry.publish_chunk(task_id, start_chunk)
+
+            # Choose service based on configuration
+            use_sdk = config.use_claude_agent_sdk
+            stream_fn = (
+                sdk_service.stream_chat_completion_sdk
+                if use_sdk
+                else chat_service.stream_chat_completion
+            )
+            # Pass message=None since we already added it to the session above
+            async for chunk in stream_fn(
+                session_id,
+                None,  # Message already in session
+                is_user_message=request.is_user_message,
+                user_id=user_id,
+                session=session,  # Pass session with message already added
+                context=request.context,
+            ):
+                chunk_count += 1
+                # Write to Redis (subscribers will receive via XREAD)
+                await stream_registry.publish_chunk(task_id, chunk)
+
+            logger.info(
+                f"[BG_TASK] AI generation completed for session {session_id}: {chunk_count} chunks, marking task {task_id} as completed"
+            )
+            # Mark task as completed (also publishes StreamFinish)
+            completed = await stream_registry.mark_task_completed(task_id, "completed")
+            logger.info(f"[BG_TASK] mark_task_completed returned: {completed}")
+        except Exception as e:
+            logger.error(
+                f"Error in background AI generation for session {session_id}: {e}"
+            )
+            await stream_registry.mark_task_completed(task_id, "failed")
+
+    # Start the AI generation in a background task
+    bg_task = asyncio.create_task(run_ai_generation())
+    await stream_registry.set_task_asyncio_task(task_id, bg_task)
+
+    # SSE endpoint that subscribes to the task's stream
+    async def event_generator() -> AsyncGenerator[str, None]:
+        subscriber_queue = None
+        try:
+            # Subscribe to the task stream (replays + live updates)
+            subscriber_queue = await stream_registry.subscribe_to_task(
+                task_id=task_id,
+                user_id=user_id,
+                last_message_id="0-0",  # Get all messages from the beginning
+            )
+
+            if subscriber_queue is None:
+                logger.warning(f"Failed to subscribe to task {task_id}")
+                yield StreamFinish().to_sse()
+                yield "data: [DONE]\n\n"
+                return
+
+            # Read from the subscriber queue and yield to SSE
+            while True:
+                try:
+                    chunk = await asyncio.wait_for(subscriber_queue.get(), timeout=30.0)
+                    yield chunk.to_sse()
+
+                    # Check for finish signal
+                    if isinstance(chunk, StreamFinish):
+                        break
+                except asyncio.TimeoutError:
+                    # Send heartbeat to keep connection alive
+                    yield StreamHeartbeat().to_sse()
+
+        except GeneratorExit:
+            pass  # Client disconnected - normal behavior
+        except Exception as e:
+            logger.error(f"Error in SSE stream for task {task_id}: {e}")
+        finally:
+            # Unsubscribe when client disconnects or stream ends
+            if subscriber_queue is not None:
+                try:
+                    await stream_registry.unsubscribe_from_task(
+                        task_id, subscriber_queue
+                    )
+                except Exception as unsub_err:
+                    logger.error(
+                        f"Error unsubscribing from task {task_id}: {unsub_err}",
+                        exc_info=True,
+                    )
+            # AI SDK protocol termination - always yield even if unsubscribe fails
+            yield "data: [DONE]\n\n"
 
     return StreamingResponse(
         event_generator(),
@@ -296,35 +449,21 @@ async def stream_chat_get(
     session = await _validate_and_get_session(session_id, user_id)
 
     async def event_generator() -> AsyncGenerator[str, None]:
-        chunk_count = 0
+        # Choose service based on configuration
-        first_chunk_type: str | None = None
+        use_sdk = config.use_claude_agent_sdk
-        async for chunk in chat_service.stream_chat_completion(
+        stream_fn = (
+            sdk_service.stream_chat_completion_sdk
+            if use_sdk
+            else chat_service.stream_chat_completion
+        )
+        async for chunk in stream_fn(
             session_id,
             message,
             is_user_message=is_user_message,
             user_id=user_id,
             session=session,  # Pass pre-fetched session to avoid double-fetch
         ):
-            if chunk_count < 3:
-                logger.info(
-                    "Chat stream chunk",
-                    extra={
-                        "session_id": session_id,
-                        "chunk_type": str(chunk.type),
-                    },
-                )
-            if not first_chunk_type:
-                first_chunk_type = str(chunk.type)
-            chunk_count += 1
             yield chunk.to_sse()
-        logger.info(
-            "Chat stream completed",
-            extra={
-                "session_id": session_id,
-                "chunk_count": chunk_count,
-                "first_chunk_type": first_chunk_type,
-            },
-        )
         # AI SDK protocol termination
         yield "data: [DONE]\n\n"
 
@@ -366,6 +505,249 @@ async def session_assign_user(
     return {"status": "ok"}
 
 
+# ========== Task Streaming (SSE Reconnection) ==========
+
+
+@router.get(
+    "/tasks/{task_id}/stream",
+)
+async def stream_task(
+    task_id: str,
+    user_id: str | None = Depends(auth.get_user_id),
+    last_message_id: str = Query(
+        default="0-0",
+        description="Last Redis Stream message ID received (e.g., '1706540123456-0'). Use '0-0' for full replay.",
+    ),
+):
+    """
+    Reconnect to a long-running task's SSE stream.
+
+    When a long-running operation (like agent generation) starts, the client
+    receives a task_id. If the connection drops, the client can reconnect
+    using this endpoint to resume receiving updates.
+
+    Args:
+        task_id: The task ID from the operation_started response.
+        user_id: Authenticated user ID for ownership validation.
+        last_message_id: Last Redis Stream message ID received ("0-0" for full replay).
+
+    Returns:
+        StreamingResponse: SSE-formatted response chunks starting after last_message_id.
+
+    Raises:
+        HTTPException: 404 if task not found, 410 if task expired, 403 if access denied.
+    """
+    # Check task existence and expiry before subscribing
+    task, error_code = await stream_registry.get_task_with_expiry_info(task_id)
+
+    if error_code == "TASK_EXPIRED":
+        raise HTTPException(
+            status_code=410,
+            detail={
+                "code": "TASK_EXPIRED",
+                "message": "This operation has expired. Please try again.",
+            },
+        )
+
+    if error_code == "TASK_NOT_FOUND":
+        raise HTTPException(
+            status_code=404,
+            detail={
+                "code": "TASK_NOT_FOUND",
+                "message": f"Task {task_id} not found.",
+            },
+        )
+
+    # Validate ownership if task has an owner
+    if task and task.user_id and user_id != task.user_id:
+        raise HTTPException(
+            status_code=403,
+            detail={
+                "code": "ACCESS_DENIED",
+                "message": "You do not have access to this task.",
+            },
+        )
+
+    # Get subscriber queue from stream registry
+    subscriber_queue = await stream_registry.subscribe_to_task(
+        task_id=task_id,
+        user_id=user_id,
+        last_message_id=last_message_id,
+    )
+
+    if subscriber_queue is None:
+        raise HTTPException(
+            status_code=404,
+            detail={
+                "code": "TASK_NOT_FOUND",
+                "message": f"Task {task_id} not found or access denied.",
+            },
+        )
+
+    async def event_generator() -> AsyncGenerator[str, None]:
+        heartbeat_interval = 15.0  # Send heartbeat every 15 seconds
+        try:
+            while True:
+                try:
+                    # Wait for next chunk with timeout for heartbeats
+                    chunk = await asyncio.wait_for(
+                        subscriber_queue.get(), timeout=heartbeat_interval
+                    )
+                    yield chunk.to_sse()
+
+                    # Check for finish signal
+                    if isinstance(chunk, StreamFinish):
+                        break
+                except asyncio.TimeoutError:
+                    # Send heartbeat to keep connection alive
+                    yield StreamHeartbeat().to_sse()
+        except Exception as e:
+            logger.error(f"Error in task stream {task_id}: {e}", exc_info=True)
+        finally:
+            # Unsubscribe when client disconnects or stream ends
+            try:
+                await stream_registry.unsubscribe_from_task(task_id, subscriber_queue)
+            except Exception as unsub_err:
+                logger.error(
+                    f"Error unsubscribing from task {task_id}: {unsub_err}",
+                    exc_info=True,
+                )
+            # AI SDK protocol termination - always yield even if unsubscribe fails
+            yield "data: [DONE]\n\n"
+
+    return StreamingResponse(
+        event_generator(),
+        media_type="text/event-stream",
+        headers={
+            "Cache-Control": "no-cache",
+            "Connection": "keep-alive",
+            "X-Accel-Buffering": "no",
+            "x-vercel-ai-ui-message-stream": "v1",
+        },
+    )
+
+
+@router.get(
+    "/tasks/{task_id}",
+)
+async def get_task_status(
+    task_id: str,
+    user_id: str | None = Depends(auth.get_user_id),
+) -> dict:
+    """
+    Get the status of a long-running task.
+
+    Args:
+        task_id: The task ID to check.
+        user_id: Authenticated user ID for ownership validation.
+
+    Returns:
+        dict: Task status including task_id, status, tool_name, and operation_id.
+
+    Raises:
+        NotFoundError: If task_id is not found or user doesn't have access.
+    """
+    task = await stream_registry.get_task(task_id)
+
+    if task is None:
+        raise NotFoundError(f"Task {task_id} not found.")
+
+    # Validate ownership - if task has an owner, requester must match
+    if task.user_id and user_id != task.user_id:
+        raise NotFoundError(f"Task {task_id} not found.")
+
+    return {
+        "task_id": task.task_id,
+        "session_id": task.session_id,
+        "status": task.status,
+        "tool_name": task.tool_name,
+        "operation_id": task.operation_id,
+        "created_at": task.created_at.isoformat(),
+    }
+
+
+# ========== External Completion Webhook ==========
+
+
+@router.post(
+    "/operations/{operation_id}/complete",
+    status_code=200,
+)
+async def complete_operation(
+    operation_id: str,
+    request: OperationCompleteRequest,
+    x_api_key: str | None = Header(default=None),
+) -> dict:
+    """
+    External completion webhook for long-running operations.
+
+    Called by Agent Generator (or other services) when an operation completes.
+    This triggers the stream registry to publish completion and continue LLM generation.
+
+    Args:
+        operation_id: The operation ID to complete.
+        request: Completion payload with success status and result/error.
+        x_api_key: Internal API key for authentication.
+
+    Returns:
+        dict: Status of the completion.
+
+    Raises:
+        HTTPException: If API key is invalid or operation not found.
+    """
+    # Validate internal API key - reject if not configured or invalid
+    if not config.internal_api_key:
+        logger.error(
+            "Operation complete webhook rejected: CHAT_INTERNAL_API_KEY not configured"
+        )
+        raise HTTPException(
+            status_code=503,
+            detail="Webhook not available: internal API key not configured",
+        )
+    if x_api_key != config.internal_api_key:
+        raise HTTPException(status_code=401, detail="Invalid API key")
+
+    # Find task by operation_id
+    task = await stream_registry.find_task_by_operation_id(operation_id)
+    if task is None:
+        raise HTTPException(
+            status_code=404,
+            detail=f"Operation {operation_id} not found",
+        )
+
+    logger.info(
+        f"Received completion webhook for operation {operation_id} "
+        f"(task_id={task.task_id}, success={request.success})"
+    )
+
+    if request.success:
+        await process_operation_success(task, request.result)
+    else:
+        await process_operation_failure(task, request.error)
+
+    return {"status": "ok", "task_id": task.task_id}
+
+
+# ========== Configuration ==========
+
+
+@router.get("/config/ttl", status_code=200)
+async def get_ttl_config() -> dict:
+    """
+    Get the stream TTL configuration.
+
+    Returns the Time-To-Live settings for chat streams, which determines
+    how long clients can reconnect to an active stream.
+
+    Returns:
+        dict: TTL configuration with seconds and milliseconds values.
+    """
+    return {
+        "stream_ttl_seconds": config.stream_ttl,
+        "stream_ttl_ms": config.stream_ttl * 1000,
+    }
+
+
 # ========== Health Check ==========
 
 

autogpt_platform/backend/backend/api/features/chat/sdk/__init__.py

@@ -0,0 +1,14 @@
+"""Claude Agent SDK integration for CoPilot.
+
+This module provides the integration layer between the Claude Agent SDK
+and the existing CoPilot tool system, enabling drop-in replacement of
+the current LLM orchestration with the battle-tested Claude Agent SDK.
+"""
+
+from .service import stream_chat_completion_sdk
+from .tool_adapter import create_copilot_mcp_server
+
+__all__ = [
+    "stream_chat_completion_sdk",
+    "create_copilot_mcp_server",
+]

autogpt_platform/backend/backend/api/features/chat/sdk/anthropic_fallback.py

@@ -0,0 +1,348 @@
+"""Anthropic SDK fallback implementation.
+
+This module provides the fallback streaming implementation using the Anthropic SDK
+directly when the Claude Agent SDK is not available.
+"""
+
+import json
+import logging
+import os
+import uuid
+from collections.abc import AsyncGenerator
+from typing import Any, cast
+
+from ..model import ChatMessage, ChatSession
+from ..response_model import (
+    StreamBaseResponse,
+    StreamError,
+    StreamFinish,
+    StreamTextDelta,
+    StreamTextEnd,
+    StreamTextStart,
+    StreamToolInputAvailable,
+    StreamToolInputStart,
+    StreamToolOutputAvailable,
+    StreamUsage,
+)
+from .tool_adapter import get_tool_definitions, get_tool_handlers
+
+logger = logging.getLogger(__name__)
+
+
+async def stream_with_anthropic(
+    session: ChatSession,
+    system_prompt: str,
+    text_block_id: str,
+) -> AsyncGenerator[StreamBaseResponse, None]:
+    """Stream using Anthropic SDK directly with tool calling support.
+
+    This function accumulates messages into the session for persistence.
+    The caller should NOT yield an additional StreamFinish - this function handles it.
+    """
+    import anthropic
+
+    # Only use ANTHROPIC_API_KEY - don't fall back to OpenRouter keys
+    api_key = os.getenv("ANTHROPIC_API_KEY")
+    if not api_key:
+        yield StreamError(
+            errorText="ANTHROPIC_API_KEY not configured for fallback",
+            code="config_error",
+        )
+        yield StreamFinish()
+        return
+
+    client = anthropic.AsyncAnthropic(api_key=api_key)
+    tool_definitions = get_tool_definitions()
+    tool_handlers = get_tool_handlers()
+
+    anthropic_tools = [
+        {
+            "name": t["name"],
+            "description": t["description"],
+            "input_schema": t["inputSchema"],
+        }
+        for t in tool_definitions
+    ]
+
+    anthropic_messages = _convert_session_to_anthropic(session)
+
+    if not anthropic_messages or anthropic_messages[-1]["role"] != "user":
+        anthropic_messages.append(
+            {"role": "user", "content": "Continue with the task."}
+        )
+
+    has_started_text = False
+    max_iterations = 10
+    accumulated_text = ""
+    accumulated_tool_calls: list[dict[str, Any]] = []
+
+    for _ in range(max_iterations):
+        try:
+            async with client.messages.stream(
+                model="claude-sonnet-4-20250514",
+                max_tokens=4096,
+                system=system_prompt,
+                messages=cast(Any, anthropic_messages),
+                tools=cast(Any, anthropic_tools) if anthropic_tools else [],
+            ) as stream:
+                async for event in stream:
+                    if event.type == "content_block_start":
+                        block = event.content_block
+                        if hasattr(block, "type"):
+                            if block.type == "text" and not has_started_text:
+                                yield StreamTextStart(id=text_block_id)
+                                has_started_text = True
+                            elif block.type == "tool_use":
+                                yield StreamToolInputStart(
+                                    toolCallId=block.id, toolName=block.name
+                                )
+
+                    elif event.type == "content_block_delta":
+                        delta = event.delta
+                        if hasattr(delta, "type") and delta.type == "text_delta":
+                            accumulated_text += delta.text
+                            yield StreamTextDelta(id=text_block_id, delta=delta.text)
+
+                final_message = await stream.get_final_message()
+
+                if final_message.stop_reason == "tool_use":
+                    if has_started_text:
+                        yield StreamTextEnd(id=text_block_id)
+                        has_started_text = False
+                        text_block_id = str(uuid.uuid4())
+
+                    tool_results = []
+                    assistant_content: list[dict[str, Any]] = []
+
+                    for block in final_message.content:
+                        if block.type == "text":
+                            assistant_content.append(
+                                {"type": "text", "text": block.text}
+                            )
+                        elif block.type == "tool_use":
+                            assistant_content.append(
+                                {
+                                    "type": "tool_use",
+                                    "id": block.id,
+                                    "name": block.name,
+                                    "input": block.input,
+                                }
+                            )
+
+                            # Track tool call for session persistence
+                            accumulated_tool_calls.append(
+                                {
+                                    "id": block.id,
+                                    "type": "function",
+                                    "function": {
+                                        "name": block.name,
+                                        "arguments": json.dumps(
+                                            block.input
+                                            if isinstance(block.input, dict)
+                                            else {}
+                                        ),
+                                    },
+                                }
+                            )
+
+                            yield StreamToolInputAvailable(
+                                toolCallId=block.id,
+                                toolName=block.name,
+                                input=(
+                                    block.input if isinstance(block.input, dict) else {}
+                                ),
+                            )
+
+                            output, is_error = await _execute_tool(
+                                block.name, block.input, tool_handlers
+                            )
+
+                            yield StreamToolOutputAvailable(
+                                toolCallId=block.id,
+                                toolName=block.name,
+                                output=output,
+                                success=not is_error,
+                            )
+
+                            # Save tool result to session
+                            session.messages.append(
+                                ChatMessage(
+                                    role="tool",
+                                    content=output,
+                                    tool_call_id=block.id,
+                                )
+                            )
+
+                            tool_results.append(
+                                {
+                                    "type": "tool_result",
+                                    "tool_use_id": block.id,
+                                    "content": output,
+                                    "is_error": is_error,
+                                }
+                            )
+
+                    # Save assistant message with tool calls to session
+                    session.messages.append(
+                        ChatMessage(
+                            role="assistant",
+                            content=accumulated_text or None,
+                            tool_calls=(
+                                accumulated_tool_calls
+                                if accumulated_tool_calls
+                                else None
+                            ),
+                        )
+                    )
+                    # Reset for next iteration
+                    accumulated_text = ""
+                    accumulated_tool_calls = []
+
+                    anthropic_messages.append(
+                        {"role": "assistant", "content": assistant_content}
+                    )
+                    anthropic_messages.append({"role": "user", "content": tool_results})
+                    continue
+
+                else:
+                    if has_started_text:
+                        yield StreamTextEnd(id=text_block_id)
+
+                    # Save final assistant response to session
+                    if accumulated_text:
+                        session.messages.append(
+                            ChatMessage(role="assistant", content=accumulated_text)
+                        )
+
+                    yield StreamUsage(
+                        promptTokens=final_message.usage.input_tokens,
+                        completionTokens=final_message.usage.output_tokens,
+                        totalTokens=final_message.usage.input_tokens
+                        + final_message.usage.output_tokens,
+                    )
+                    yield StreamFinish()
+                    return
+
+        except Exception as e:
+            logger.error(f"[Anthropic Fallback] Error: {e}", exc_info=True)
+            yield StreamError(
+                errorText="An error occurred. Please try again.",
+                code="anthropic_error",
+            )
+            yield StreamFinish()
+            return
+
+    yield StreamError(errorText="Max tool iterations reached", code="max_iterations")
+    yield StreamFinish()
+
+
+def _convert_session_to_anthropic(session: ChatSession) -> list[dict[str, Any]]:
+    """Convert session messages to Anthropic format.
+
+    Handles merging consecutive same-role messages (Anthropic requires alternating roles).
+    """
+    messages: list[dict[str, Any]] = []
+
+    for msg in session.messages:
+        if msg.role == "user":
+            new_msg = {"role": "user", "content": msg.content or ""}
+        elif msg.role == "assistant":
+            content: list[dict[str, Any]] = []
+            if msg.content:
+                content.append({"type": "text", "text": msg.content})
+            if msg.tool_calls:
+                for tc in msg.tool_calls:
+                    func = tc.get("function", {})
+                    args = func.get("arguments", {})
+                    if isinstance(args, str):
+                        try:
+                            args = json.loads(args)
+                        except json.JSONDecodeError:
+                            args = {}
+                    content.append(
+                        {
+                            "type": "tool_use",
+                            "id": tc.get("id", str(uuid.uuid4())),
+                            "name": func.get("name", ""),
+                            "input": args,
+                        }
+                    )
+            if content:
+                new_msg = {"role": "assistant", "content": content}
+            else:
+                continue  # Skip empty assistant messages
+        elif msg.role == "tool":
+            new_msg = {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": msg.tool_call_id or "",
+                        "content": msg.content or "",
+                    }
+                ],
+            }
+        else:
+            continue
+
+        messages.append(new_msg)
+
+    # Merge consecutive same-role messages (Anthropic requires alternating roles)
+    return _merge_consecutive_roles(messages)
+
+
+def _merge_consecutive_roles(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    """Merge consecutive messages with the same role.
+
+    Anthropic API requires alternating user/assistant roles.
+    """
+    if not messages:
+        return []
+
+    merged: list[dict[str, Any]] = []
+    for msg in messages:
+        if merged and merged[-1]["role"] == msg["role"]:
+            # Merge with previous message
+            prev_content = merged[-1]["content"]
+            new_content = msg["content"]
+
+            # Normalize both to list-of-blocks form
+            if isinstance(prev_content, str):
+                prev_content = [{"type": "text", "text": prev_content}]
+            if isinstance(new_content, str):
+                new_content = [{"type": "text", "text": new_content}]
+
+            # Ensure both are lists
+            if not isinstance(prev_content, list):
+                prev_content = [prev_content]
+            if not isinstance(new_content, list):
+                new_content = [new_content]
+
+            merged[-1]["content"] = prev_content + new_content
+        else:
+            merged.append(msg)
+
+    return merged
+
+
+async def _execute_tool(
+    tool_name: str, tool_input: Any, handlers: dict[str, Any]
+) -> tuple[str, bool]:
+    """Execute a tool and return (output, is_error)."""
+    handler = handlers.get(tool_name)
+    if not handler:
+        return f"Unknown tool: {tool_name}", True
+
+    try:
+        result = await handler(tool_input)
+        # Safely extract output - handle empty or missing content
+        content = result.get("content") or []
+        if content and isinstance(content, list) and len(content) > 0:
+            first_item = content[0]
+            output = first_item.get("text", "") if isinstance(first_item, dict) else ""
+        else:
+            output = ""
+        is_error = result.get("isError", False)
+        return output, is_error
+    except Exception as e:
+        return f"Error: {str(e)}", True

autogpt_platform/backend/backend/api/features/chat/sdk/response_adapter.py

@@ -0,0 +1,300 @@
+"""Response adapter for converting Claude Agent SDK messages to Vercel AI SDK format.
+
+This module provides the adapter layer that converts streaming messages from
+the Claude Agent SDK into the Vercel AI SDK UI Stream Protocol format that
+the frontend expects.
+"""
+
+import json
+import logging
+import uuid
+from typing import Any, AsyncGenerator
+
+from backend.api.features.chat.response_model import (
+    StreamBaseResponse,
+    StreamError,
+    StreamFinish,
+    StreamHeartbeat,
+    StreamStart,
+    StreamTextDelta,
+    StreamTextEnd,
+    StreamTextStart,
+    StreamToolInputAvailable,
+    StreamToolInputStart,
+    StreamToolOutputAvailable,
+    StreamUsage,
+)
+
+logger = logging.getLogger(__name__)
+
+
+class SDKResponseAdapter:
+    """Adapter for converting Claude Agent SDK messages to Vercel AI SDK format.
+
+    This class maintains state during a streaming session to properly track
+    text blocks, tool calls, and message lifecycle.
+    """
+
+    def __init__(self, message_id: str | None = None):
+        """Initialize the adapter.
+
+        Args:
+            message_id: Optional message ID. If not provided, one will be generated.
+        """
+        self.message_id = message_id or str(uuid.uuid4())
+        self.text_block_id = str(uuid.uuid4())
+        self.has_started_text = False
+        self.has_ended_text = False
+        self.current_tool_calls: dict[str, dict[str, Any]] = {}
+        self.task_id: str | None = None
+
+    def set_task_id(self, task_id: str) -> None:
+        """Set the task ID for reconnection support."""
+        self.task_id = task_id
+
+    def convert_message(self, sdk_message: Any) -> list[StreamBaseResponse]:
+        """Convert a single SDK message to Vercel AI SDK format.
+
+        Args:
+            sdk_message: A message from the Claude Agent SDK.
+
+        Returns:
+            List of StreamBaseResponse objects (may be empty or multiple).
+        """
+        responses: list[StreamBaseResponse] = []
+
+        # Handle different SDK message types - use class name since SDK uses dataclasses
+        class_name = type(sdk_message).__name__
+        msg_subtype = getattr(sdk_message, "subtype", None)
+
+        if class_name == "SystemMessage":
+            if msg_subtype == "init":
+                # Session initialization - emit start
+                responses.append(
+                    StreamStart(
+                        messageId=self.message_id,
+                        taskId=self.task_id,
+                    )
+                )
+
+        elif class_name == "AssistantMessage":
+            # Assistant message with content blocks
+            content = getattr(sdk_message, "content", [])
+            for block in content:
+                # Check block type by class name (SDK uses dataclasses) or dict type
+                block_class = type(block).__name__
+                block_type = block.get("type") if isinstance(block, dict) else None
+
+                if block_class == "TextBlock" or block_type == "text":
+                    # Text content
+                    text = getattr(block, "text", None) or (
+                        block.get("text") if isinstance(block, dict) else ""
+                    )
+
+                    if text:
+                        # Start text block if needed (or restart after tool calls)
+                        if not self.has_started_text or self.has_ended_text:
+                            # Generate new text block ID for text after tools
+                            if self.has_ended_text:
+                                self.text_block_id = str(uuid.uuid4())
+                                self.has_ended_text = False
+                            responses.append(StreamTextStart(id=self.text_block_id))
+                            self.has_started_text = True
+
+                        # Emit text delta
+                        responses.append(
+                            StreamTextDelta(
+                                id=self.text_block_id,
+                                delta=text,
+                            )
+                        )
+
+                elif block_class == "ToolUseBlock" or block_type == "tool_use":
+                    # Tool call
+                    tool_id_raw = getattr(block, "id", None) or (
+                        block.get("id") if isinstance(block, dict) else None
+                    )
+                    tool_id: str = (
+                        str(tool_id_raw) if tool_id_raw else str(uuid.uuid4())
+                    )
+
+                    tool_name_raw = getattr(block, "name", None) or (
+                        block.get("name") if isinstance(block, dict) else None
+                    )
+                    tool_name: str = str(tool_name_raw) if tool_name_raw else "unknown"
+
+                    tool_input = getattr(block, "input", None) or (
+                        block.get("input") if isinstance(block, dict) else {}
+                    )
+
+                    # End text block if we were streaming text
+                    if self.has_started_text and not self.has_ended_text:
+                        responses.append(StreamTextEnd(id=self.text_block_id))
+                        self.has_ended_text = True
+
+                    # Emit tool input start
+                    responses.append(
+                        StreamToolInputStart(
+                            toolCallId=tool_id,
+                            toolName=tool_name,
+                        )
+                    )
+
+                    # Emit tool input available with full input
+                    responses.append(
+                        StreamToolInputAvailable(
+                            toolCallId=tool_id,
+                            toolName=tool_name,
+                            input=tool_input if isinstance(tool_input, dict) else {},
+                        )
+                    )
+
+                    # Track the tool call
+                    self.current_tool_calls[tool_id] = {
+                        "name": tool_name,
+                        "input": tool_input,
+                    }
+
+        elif class_name in ("ToolResultMessage", "UserMessage"):
+            # Tool result - check for tool_result content
+            content = getattr(sdk_message, "content", [])
+
+            for block in content:
+                block_class = type(block).__name__
+                block_type = block.get("type") if isinstance(block, dict) else None
+
+                if block_class == "ToolResultBlock" or block_type == "tool_result":
+                    tool_use_id = getattr(block, "tool_use_id", None) or (
+                        block.get("tool_use_id") if isinstance(block, dict) else None
+                    )
+                    result_content = getattr(block, "content", None) or (
+                        block.get("content") if isinstance(block, dict) else ""
+                    )
+                    is_error = getattr(block, "is_error", False) or (
+                        block.get("is_error", False)
+                        if isinstance(block, dict)
+                        else False
+                    )
+
+                    if tool_use_id:
+                        tool_info = self.current_tool_calls.get(tool_use_id, {})
+                        tool_name = tool_info.get("name", "unknown")
+
+                        # Format the output
+                        if isinstance(result_content, list):
+                            # Extract text from content blocks
+                            output_text = ""
+                            for item in result_content:
+                                if (
+                                    isinstance(item, dict)
+                                    and item.get("type") == "text"
+                                ):
+                                    output_text += item.get("text", "")
+                                elif hasattr(item, "text"):
+                                    output_text += getattr(item, "text", "")
+                            output = output_text
+                        elif isinstance(result_content, str):
+                            output = result_content
+                        else:
+                            output = json.dumps(result_content)
+
+                        responses.append(
+                            StreamToolOutputAvailable(
+                                toolCallId=tool_use_id,
+                                toolName=tool_name,
+                                output=output,
+                                success=not is_error,
+                            )
+                        )
+
+        elif class_name == "ResultMessage":
+            # Final result
+            if msg_subtype == "success":
+                # End text block if still open
+                if self.has_started_text and not self.has_ended_text:
+                    responses.append(StreamTextEnd(id=self.text_block_id))
+                    self.has_ended_text = True
+
+                # Emit finish
+                responses.append(StreamFinish())
+
+            elif msg_subtype in ("error", "error_during_execution"):
+                error_msg = getattr(sdk_message, "error", "Unknown error")
+                responses.append(
+                    StreamError(
+                        errorText=str(error_msg),
+                        code="sdk_error",
+                    )
+                )
+                responses.append(StreamFinish())
+
+        elif class_name == "ErrorMessage":
+            # Error message
+            error_msg = getattr(sdk_message, "message", None) or getattr(
+                sdk_message, "error", "Unknown error"
+            )
+            responses.append(
+                StreamError(
+                    errorText=str(error_msg),
+                    code="sdk_error",
+                )
+            )
+            responses.append(StreamFinish())
+
+        return responses
+
+    def create_heartbeat(self, tool_call_id: str | None = None) -> StreamHeartbeat:
+        """Create a heartbeat response."""
+        return StreamHeartbeat(toolCallId=tool_call_id)
+
+    def create_usage(
+        self,
+        prompt_tokens: int,
+        completion_tokens: int,
+    ) -> StreamUsage:
+        """Create a usage statistics response."""
+        return StreamUsage(
+            promptTokens=prompt_tokens,
+            completionTokens=completion_tokens,
+            totalTokens=prompt_tokens + completion_tokens,
+        )
+
+
+async def adapt_sdk_stream(
+    sdk_stream: AsyncGenerator[Any, None],
+    message_id: str | None = None,
+    task_id: str | None = None,
+) -> AsyncGenerator[StreamBaseResponse, None]:
+    """Adapt a Claude Agent SDK stream to Vercel AI SDK format.
+
+    Args:
+        sdk_stream: The async generator from the Claude Agent SDK.
+        message_id: Optional message ID for the response.
+        task_id: Optional task ID for reconnection support.
+
+    Yields:
+        StreamBaseResponse objects in Vercel AI SDK format.
+    """
+    adapter = SDKResponseAdapter(message_id=message_id)
+    if task_id:
+        adapter.set_task_id(task_id)
+
+    # Emit start immediately
+    yield StreamStart(messageId=adapter.message_id, taskId=task_id)
+
+    try:
+        async for sdk_message in sdk_stream:
+            responses = adapter.convert_message(sdk_message)
+            for response in responses:
+                # Skip duplicate start messages
+                if isinstance(response, StreamStart):
+                    continue
+                yield response
+
+    except Exception as e:
+        logger.error(f"Error in SDK stream: {e}", exc_info=True)
+        yield StreamError(
+            errorText=f"Stream error: {str(e)}",
+            code="stream_error",
+        )
+        yield StreamFinish()

autogpt_platform/backend/backend/api/features/chat/sdk/security_hooks.py

@@ -0,0 +1,278 @@
+"""Security hooks for Claude Agent SDK integration.
+
+This module provides security hooks that validate tool calls before execution,
+ensuring multi-user isolation and preventing unauthorized operations.
+"""
+
+import logging
+import re
+from typing import Any, cast
+
+logger = logging.getLogger(__name__)
+
+# Tools that are blocked entirely (CLI/system access)
+BLOCKED_TOOLS = {
+    "Bash",
+    "bash",
+    "shell",
+    "exec",
+    "terminal",
+    "command",
+    "Read",  # Block raw file read - use workspace tools instead
+    "Write",  # Block raw file write - use workspace tools instead
+    "Edit",  # Block raw file edit - use workspace tools instead
+    "Glob",  # Block raw file glob - use workspace tools instead
+    "Grep",  # Block raw file grep - use workspace tools instead
+}
+
+# Dangerous patterns in tool inputs
+DANGEROUS_PATTERNS = [
+    r"sudo",
+    r"rm\s+-rf",
+    r"dd\s+if=",
+    r"/etc/passwd",
+    r"/etc/shadow",
+    r"chmod\s+777",
+    r"curl\s+.*\|.*sh",
+    r"wget\s+.*\|.*sh",
+    r"eval\s*\(",
+    r"exec\s*\(",
+    r"__import__",
+    r"os\.system",
+    r"subprocess",
+]
+
+
+def _validate_tool_access(tool_name: str, tool_input: dict[str, Any]) -> dict[str, Any]:
+    """Validate that a tool call is allowed.
+
+    Returns:
+        Empty dict to allow, or dict with hookSpecificOutput to deny
+    """
+    # Block forbidden tools
+    if tool_name in BLOCKED_TOOLS:
+        logger.warning(f"Blocked tool access attempt: {tool_name}")
+        return {
+            "hookSpecificOutput": {
+                "hookEventName": "PreToolUse",
+                "permissionDecision": "deny",
+                "permissionDecisionReason": (
+                    f"Tool '{tool_name}' is not available. "
+                    "Use the CoPilot-specific tools instead."
+                ),
+            }
+        }
+
+    # Check for dangerous patterns in tool input
+    input_str = str(tool_input)
+
+    for pattern in DANGEROUS_PATTERNS:
+        if re.search(pattern, input_str, re.IGNORECASE):
+            logger.warning(
+                f"Blocked dangerous pattern in tool input: {pattern} in {tool_name}"
+            )
+            return {
+                "hookSpecificOutput": {
+                    "hookEventName": "PreToolUse",
+                    "permissionDecision": "deny",
+                    "permissionDecisionReason": "Input contains blocked pattern",
+                }
+            }
+
+    return {}
+
+
+def _validate_user_isolation(
+    tool_name: str, tool_input: dict[str, Any], user_id: str | None
+) -> dict[str, Any]:
+    """Validate that tool calls respect user isolation."""
+    # For workspace file tools, ensure path doesn't escape
+    if "workspace" in tool_name.lower():
+        path = tool_input.get("path", "") or tool_input.get("file_path", "")
+        if path:
+            # Check for path traversal
+            if ".." in path or path.startswith("/"):
+                logger.warning(
+                    f"Blocked path traversal attempt: {path} by user {user_id}"
+                )
+                return {
+                    "hookSpecificOutput": {
+                        "hookEventName": "PreToolUse",
+                        "permissionDecision": "deny",
+                        "permissionDecisionReason": "Path traversal not allowed",
+                    }
+                }
+
+    return {}
+
+
+def create_security_hooks(user_id: str | None) -> dict[str, Any]:
+    """Create the security hooks configuration for Claude Agent SDK.
+
+    Includes security validation and observability hooks:
+    - PreToolUse: Security validation before tool execution
+    - PostToolUse: Log successful tool executions
+    - PostToolUseFailure: Log and handle failed tool executions
+    - PreCompact: Log context compaction events (SDK handles compaction automatically)
+
+    Args:
+        user_id: Current user ID for isolation validation
+
+    Returns:
+        Hooks configuration dict for ClaudeAgentOptions
+    """
+    try:
+        from claude_agent_sdk import HookMatcher
+        from claude_agent_sdk.types import HookContext, HookInput, SyncHookJSONOutput
+
+        async def pre_tool_use_hook(
+            input_data: HookInput,
+            tool_use_id: str | None,
+            context: HookContext,
+        ) -> SyncHookJSONOutput:
+            """Combined pre-tool-use validation hook."""
+            _ = context  # unused but required by signature
+            tool_name = cast(str, input_data.get("tool_name", ""))
+            tool_input = cast(dict[str, Any], input_data.get("tool_input", {}))
+
+            # Validate basic tool access
+            result = _validate_tool_access(tool_name, tool_input)
+            if result:
+                return cast(SyncHookJSONOutput, result)
+
+            # Validate user isolation
+            result = _validate_user_isolation(tool_name, tool_input, user_id)
+            if result:
+                return cast(SyncHookJSONOutput, result)
+
+            logger.debug(f"[SDK] Tool start: {tool_name}, user={user_id}")
+            return cast(SyncHookJSONOutput, {})
+
+        async def post_tool_use_hook(
+            input_data: HookInput,
+            tool_use_id: str | None,
+            context: HookContext,
+        ) -> SyncHookJSONOutput:
+            """Log successful tool executions for observability."""
+            _ = context
+            tool_name = cast(str, input_data.get("tool_name", ""))
+            logger.debug(f"[SDK] Tool success: {tool_name}, tool_use_id={tool_use_id}")
+            return cast(SyncHookJSONOutput, {})
+
+        async def post_tool_failure_hook(
+            input_data: HookInput,
+            tool_use_id: str | None,
+            context: HookContext,
+        ) -> SyncHookJSONOutput:
+            """Log failed tool executions for debugging."""
+            _ = context
+            tool_name = cast(str, input_data.get("tool_name", ""))
+            error = input_data.get("error", "Unknown error")
+            logger.warning(
+                f"[SDK] Tool failed: {tool_name}, error={error}, "
+                f"user={user_id}, tool_use_id={tool_use_id}"
+            )
+            return cast(SyncHookJSONOutput, {})
+
+        async def pre_compact_hook(
+            input_data: HookInput,
+            tool_use_id: str | None,
+            context: HookContext,
+        ) -> SyncHookJSONOutput:
+            """Log when SDK triggers context compaction.
+
+            The SDK automatically compacts conversation history when it grows too large.
+            This hook provides visibility into when compaction happens.
+            """
+            _ = context, tool_use_id
+            trigger = input_data.get("trigger", "auto")
+            logger.info(
+                f"[SDK] Context compaction triggered: {trigger}, user={user_id}"
+            )
+            return cast(SyncHookJSONOutput, {})
+
+        return {
+            "PreToolUse": [HookMatcher(matcher="*", hooks=[pre_tool_use_hook])],
+            "PostToolUse": [HookMatcher(matcher="*", hooks=[post_tool_use_hook])],
+            "PostToolUseFailure": [
+                HookMatcher(matcher="*", hooks=[post_tool_failure_hook])
+            ],
+            "PreCompact": [HookMatcher(matcher="*", hooks=[pre_compact_hook])],
+        }
+    except ImportError:
+        # Fallback for when SDK isn't available - return empty hooks
+        return {}
+
+
+def create_strict_security_hooks(
+    user_id: str | None,
+    allowed_tools: list[str] | None = None,
+) -> dict[str, Any]:
+    """Create strict security hooks that only allow specific tools.
+
+    Args:
+        user_id: Current user ID
+        allowed_tools: List of allowed tool names (defaults to CoPilot tools)
+
+    Returns:
+        Hooks configuration dict
+    """
+    try:
+        from claude_agent_sdk import HookMatcher
+        from claude_agent_sdk.types import HookContext, HookInput, SyncHookJSONOutput
+
+        from .tool_adapter import RAW_TOOL_NAMES
+
+        tools_list = allowed_tools if allowed_tools is not None else RAW_TOOL_NAMES
+        allowed_set = set(tools_list)
+
+        async def strict_pre_tool_use(
+            input_data: HookInput,
+            tool_use_id: str | None,
+            context: HookContext,
+        ) -> SyncHookJSONOutput:
+            """Strict validation that only allows whitelisted tools."""
+            _ = context  # unused but required by signature
+            tool_name = cast(str, input_data.get("tool_name", ""))
+            tool_input = cast(dict[str, Any], input_data.get("tool_input", {}))
+
+            # Remove MCP prefix if present
+            clean_name = tool_name.removeprefix("mcp__copilot__")
+
+            if clean_name not in allowed_set:
+                logger.warning(f"Blocked non-whitelisted tool: {tool_name}")
+                return cast(
+                    SyncHookJSONOutput,
+                    {
+                        "hookSpecificOutput": {
+                            "hookEventName": "PreToolUse",
+                            "permissionDecision": "deny",
+                            "permissionDecisionReason": (
+                                f"Tool '{tool_name}' is not in the allowed list"
+                            ),
+                        }
+                    },
+                )
+
+            # Run standard validations
+            result = _validate_tool_access(tool_name, tool_input)
+            if result:
+                return cast(SyncHookJSONOutput, result)
+
+            result = _validate_user_isolation(tool_name, tool_input, user_id)
+            if result:
+                return cast(SyncHookJSONOutput, result)
+
+            logger.debug(
+                f"[SDK Audit] Tool call: tool={tool_name}, "
+                f"user={user_id}, tool_use_id={tool_use_id}"
+            )
+            return cast(SyncHookJSONOutput, {})
+
+        return {
+            "PreToolUse": [
+                HookMatcher(matcher="*", hooks=[strict_pre_tool_use]),
+            ],
+        }
+    except ImportError:
+        return {}

autogpt_platform/backend/backend/api/features/chat/sdk/service.py

@@ -0,0 +1,471 @@
+"""Claude Agent SDK service layer for CoPilot chat completions."""
+
+import asyncio
+import json
+import logging
+import uuid
+from collections.abc import AsyncGenerator
+from typing import Any
+
+import openai
+
+from backend.data.understanding import (
+    format_understanding_for_prompt,
+    get_business_understanding,
+)
+from backend.util.exceptions import NotFoundError
+
+from ..config import ChatConfig
+from ..model import (
+    ChatMessage,
+    ChatSession,
+    get_chat_session,
+    update_session_title,
+    upsert_chat_session,
+)
+from ..response_model import (
+    StreamBaseResponse,
+    StreamError,
+    StreamFinish,
+    StreamStart,
+    StreamTextDelta,
+    StreamToolInputAvailable,
+    StreamToolOutputAvailable,
+)
+from ..tracking import track_user_message
+from .anthropic_fallback import stream_with_anthropic
+from .response_adapter import SDKResponseAdapter
+from .security_hooks import create_security_hooks
+from .tool_adapter import (
+    COPILOT_TOOL_NAMES,
+    create_copilot_mcp_server,
+    set_execution_context,
+)
+
+logger = logging.getLogger(__name__)
+config = ChatConfig()
+
+# Set to hold background tasks to prevent garbage collection
+_background_tasks: set[asyncio.Task[Any]] = set()
+
+DEFAULT_SYSTEM_PROMPT = """You are **Otto**, an AI Co-Pilot for AutoGPT and a Forward-Deployed Automation Engineer serving small business owners. Your mission is to help users automate business tasks with AI by delivering tangible value through working automations—not through documentation or lengthy explanations.
+
+Here is everything you know about the current user from previous interactions:
+
+<users_information>
+{users_information}
+</users_information>
+
+## YOUR CORE MANDATE
+
+You are action-oriented. Your success is measured by:
+- **Value Delivery**: Does the user think "wow, that was amazing" or "what was the point"?
+- **Demonstrable Proof**: Show working automations, not descriptions of what's possible
+- **Time Saved**: Focus on tangible efficiency gains
+- **Quality Output**: Deliver results that meet or exceed expectations
+
+## YOUR WORKFLOW
+
+Adapt flexibly to the conversation context. Not every interaction requires all stages:
+
+1. **Explore & Understand**: Learn about the user's business, tasks, and goals. Use `add_understanding` to capture important context that will improve future conversations.
+
+2. **Assess Automation Potential**: Help the user understand whether and how AI can automate their task.
+
+3. **Prepare for AI**: Provide brief, actionable guidance on prerequisites (data, access, etc.).
+
+4. **Discover or Create Agents**:
+   - **Always check the user's library first** with `find_library_agent` (these may be customized to their needs)
+   - Search the marketplace with `find_agent` for pre-built automations
+   - Find reusable components with `find_block`
+   - Create custom solutions with `create_agent` if nothing suitable exists
+   - Modify existing library agents with `edit_agent`
+
+5. **Execute**: Run automations immediately, schedule them, or set up webhooks using `run_agent`. Test specific components with `run_block`.
+
+6. **Show Results**: Display outputs using `agent_output`.
+
+## BEHAVIORAL GUIDELINES
+
+**Be Concise:**
+- Target 2-5 short lines maximum
+- Make every word count—no repetition or filler
+- Use lightweight structure for scannability (bullets, numbered lists, short prompts)
+- Avoid jargon (blocks, slugs, cron) unless the user asks
+
+**Be Proactive:**
+- Suggest next steps before being asked
+- Anticipate needs based on conversation context and user information
+- Look for opportunities to expand scope when relevant
+- Reveal capabilities through action, not explanation
+
+**Use Tools Effectively:**
+- Select the right tool for each task
+- **Always check `find_library_agent` before searching the marketplace**
+- Use `add_understanding` to capture valuable business context
+- When tool calls fail, try alternative approaches
+
+## CRITICAL REMINDER
+
+You are NOT a chatbot. You are NOT documentation. You are a partner who helps busy business owners get value quickly by showing proof through working automations. Bias toward action over explanation."""
+
+
+async def _build_system_prompt(
+    user_id: str | None, has_conversation_history: bool = False
+) -> tuple[str, Any]:
+    """Build the system prompt with user's business understanding context.
+
+    Args:
+        user_id: The user ID to fetch understanding for.
+        has_conversation_history: Whether there's existing conversation history.
+            If True, we don't tell the model to greet/introduce (since they're
+            already in a conversation).
+    """
+    understanding = None
+    if user_id:
+        try:
+            understanding = await get_business_understanding(user_id)
+        except Exception as e:
+            logger.warning(f"Failed to fetch business understanding: {e}")
+
+    if understanding:
+        context = format_understanding_for_prompt(understanding)
+    elif has_conversation_history:
+        # Don't tell model to greet if there's conversation history
+        context = "No prior understanding saved yet. Continue the existing conversation naturally."
+    else:
+        context = "This is the first time you are meeting the user. Greet them and introduce them to the platform"
+
+    return DEFAULT_SYSTEM_PROMPT.format(users_information=context), understanding
+
+
+def _format_conversation_history(session: ChatSession) -> str:
+    """Format conversation history as a prompt context.
+
+    The SDK handles context compaction automatically, but we apply
+    max_context_messages as a safety guard to limit initial prompt size.
+    """
+    if not session.messages:
+        return ""
+
+    # Get all messages except the last user message (which will be the prompt)
+    messages = session.messages[:-1] if session.messages else []
+    if not messages:
+        return ""
+
+    # Apply max_context_messages limit as a safety guard
+    # (SDK handles compaction, but this prevents excessively large initial prompts)
+    max_messages = config.max_context_messages
+    if len(messages) > max_messages:
+        messages = messages[-max_messages:]
+
+    history_parts = ["<conversation_history>"]
+
+    for msg in messages:
+        if msg.role == "user":
+            history_parts.append(f"User: {msg.content or ''}")
+        elif msg.role == "assistant":
+            # Pass full content - SDK handles compaction automatically
+            history_parts.append(f"Assistant: {msg.content or ''}")
+            if msg.tool_calls:
+                for tc in msg.tool_calls:
+                    func = tc.get("function", {})
+                    history_parts.append(
+                        f"  [Called tool: {func.get('name', 'unknown')}]"
+                    )
+        elif msg.role == "tool":
+            # Pass full tool results - SDK handles compaction
+            history_parts.append(f"  [Tool result: {msg.content or ''}]")
+
+    history_parts.append("</conversation_history>")
+    history_parts.append("")
+    history_parts.append(
+        "Continue this conversation. Respond to the user's latest message:"
+    )
+    history_parts.append("")
+
+    return "\n".join(history_parts)
+
+
+async def _generate_session_title(
+    message: str,
+    user_id: str | None = None,
+    session_id: str | None = None,
+) -> str | None:
+    """Generate a concise title for a chat session."""
+    from backend.util.settings import Settings
+
+    settings = Settings()
+    try:
+        # Build extra_body for OpenRouter tracing
+        extra_body: dict[str, Any] = {
+            "posthogProperties": {"environment": settings.config.app_env.value},
+        }
+        if user_id:
+            extra_body["user"] = user_id[:128]
+            extra_body["posthogDistinctId"] = user_id
+        if session_id:
+            extra_body["session_id"] = session_id[:128]
+
+        client = openai.AsyncOpenAI(api_key=config.api_key, base_url=config.base_url)
+        response = await client.chat.completions.create(
+            model=config.title_model,
+            messages=[
+                {
+                    "role": "system",
+                    "content": "Generate a very short title (3-6 words) for a chat conversation based on the user's first message. Return ONLY the title, no quotes or punctuation.",
+                },
+                {"role": "user", "content": message[:500]},
+            ],
+            max_tokens=20,
+            extra_body=extra_body,
+        )
+        title = response.choices[0].message.content
+        if title:
+            title = title.strip().strip("\"'")
+            return title[:47] + "..." if len(title) > 50 else title
+        return None
+    except Exception as e:
+        logger.warning(f"Failed to generate session title: {e}")
+        return None
+
+
+async def stream_chat_completion_sdk(
+    session_id: str,
+    message: str | None = None,
+    tool_call_response: str | None = None,  # noqa: ARG001
+    is_user_message: bool = True,
+    user_id: str | None = None,
+    retry_count: int = 0,  # noqa: ARG001
+    session: ChatSession | None = None,
+    context: dict[str, str] | None = None,  # noqa: ARG001
+) -> AsyncGenerator[StreamBaseResponse, None]:
+    """Stream chat completion using Claude Agent SDK.
+
+    Drop-in replacement for stream_chat_completion with improved reliability.
+    """
+
+    if session is None:
+        session = await get_chat_session(session_id, user_id)
+
+    if not session:
+        raise NotFoundError(
+            f"Session {session_id} not found. Please create a new session first."
+        )
+
+    if message:
+        session.messages.append(
+            ChatMessage(
+                role="user" if is_user_message else "assistant", content=message
+            )
+        )
+        if is_user_message:
+            track_user_message(
+                user_id=user_id, session_id=session_id, message_length=len(message)
+            )
+
+    session = await upsert_chat_session(session)
+
+    # Generate title for new sessions (first user message)
+    if is_user_message and not session.title:
+        user_messages = [m for m in session.messages if m.role == "user"]
+        if len(user_messages) == 1:
+            first_message = user_messages[0].content or message or ""
+            if first_message:
+                task = asyncio.create_task(
+                    _update_title_async(session_id, first_message, user_id)
+                )
+                # Store reference to prevent garbage collection
+                _background_tasks.add(task)
+                task.add_done_callback(_background_tasks.discard)
+
+    # Check if there's conversation history (more than just the current message)
+    has_history = len(session.messages) > 1
+    system_prompt, _ = await _build_system_prompt(
+        user_id, has_conversation_history=has_history
+    )
+    set_execution_context(user_id, session, None)
+
+    message_id = str(uuid.uuid4())
+    text_block_id = str(uuid.uuid4())
+    task_id = str(uuid.uuid4())
+
+    yield StreamStart(messageId=message_id, taskId=task_id)
+
+    # Track whether the stream completed normally via ResultMessage
+    stream_completed = False
+
+    try:
+        try:
+            from claude_agent_sdk import ClaudeAgentOptions, ClaudeSDKClient
+
+            # Create MCP server with CoPilot tools
+            mcp_server = create_copilot_mcp_server()
+
+            options = ClaudeAgentOptions(
+                system_prompt=system_prompt,
+                mcp_servers={"copilot": mcp_server},  # type: ignore[arg-type]
+                allowed_tools=COPILOT_TOOL_NAMES,
+                hooks=create_security_hooks(user_id),  # type: ignore[arg-type]
+                continue_conversation=True,  # Enable conversation continuation
+            )
+
+            adapter = SDKResponseAdapter(message_id=message_id)
+            adapter.set_task_id(task_id)
+
+            async with ClaudeSDKClient(options=options) as client:
+                # Build prompt with conversation history for context
+                # The SDK doesn't support replaying full conversation history,
+                # so we include it as context in the prompt
+                current_message = message or ""
+                if not current_message and session.messages:
+                    last_user = [m for m in session.messages if m.role == "user"]
+                    if last_user:
+                        current_message = last_user[-1].content or ""
+
+                # Include conversation history if there are prior messages
+                if len(session.messages) > 1:
+                    history_context = _format_conversation_history(session)
+                    prompt = f"{history_context}{current_message}"
+                else:
+                    prompt = current_message
+
+                # Guard against empty prompts
+                if not prompt.strip():
+                    yield StreamError(
+                        errorText="Message cannot be empty.",
+                        code="empty_prompt",
+                    )
+                    yield StreamFinish()
+                    return
+
+                await client.query(prompt, session_id=session_id)
+
+                # Track assistant response to save to session
+                # We may need multiple assistant messages if text comes after tool results
+                assistant_response = ChatMessage(role="assistant", content="")
+                accumulated_tool_calls: list[dict[str, Any]] = []
+                has_appended_assistant = False
+                has_tool_results = False  # Track if we've received tool results
+
+                # Receive messages from the SDK
+                async for sdk_msg in client.receive_messages():
+
+                    for response in adapter.convert_message(sdk_msg):
+                        if isinstance(response, StreamStart):
+                            continue
+                        yield response
+
+                        # Accumulate text deltas into assistant response
+                        if isinstance(response, StreamTextDelta):
+                            delta = response.delta or ""
+                            # After tool results, create new assistant message for post-tool text
+                            if has_tool_results and has_appended_assistant:
+                                assistant_response = ChatMessage(
+                                    role="assistant", content=delta
+                                )
+                                accumulated_tool_calls = []  # Reset for new message
+                                session.messages.append(assistant_response)
+                                has_tool_results = False
+                            else:
+                                assistant_response.content = (
+                                    assistant_response.content or ""
+                                ) + delta
+                                if not has_appended_assistant:
+                                    session.messages.append(assistant_response)
+                                    has_appended_assistant = True
+
+                        # Track tool calls on the assistant message
+                        elif isinstance(response, StreamToolInputAvailable):
+                            accumulated_tool_calls.append(
+                                {
+                                    "id": response.toolCallId,
+                                    "type": "function",
+                                    "function": {
+                                        "name": response.toolName,
+                                        "arguments": json.dumps(response.input or {}),
+                                    },
+                                }
+                            )
+                            # Update assistant message with tool calls
+                            assistant_response.tool_calls = accumulated_tool_calls
+                            # Append assistant message if not already (tool-only response)
+                            if not has_appended_assistant:
+                                session.messages.append(assistant_response)
+                                has_appended_assistant = True
+
+                        elif isinstance(response, StreamToolOutputAvailable):
+                            session.messages.append(
+                                ChatMessage(
+                                    role="tool",
+                                    content=(
+                                        response.output
+                                        if isinstance(response.output, str)
+                                        else str(response.output)
+                                    ),
+                                    tool_call_id=response.toolCallId,
+                                )
+                            )
+                            has_tool_results = True
+
+                        elif isinstance(response, StreamFinish):
+                            stream_completed = True
+
+                    # Break out of the message loop if we received finish signal
+                    if stream_completed:
+                        break
+
+                # Ensure assistant response is saved even if no text deltas
+                # (e.g., only tool calls were made)
+                if (
+                    assistant_response.content or assistant_response.tool_calls
+                ) and not has_appended_assistant:
+                    session.messages.append(assistant_response)
+
+        except ImportError:
+            logger.warning(
+                "[SDK] claude-agent-sdk not available, using Anthropic fallback"
+            )
+            async for response in stream_with_anthropic(
+                session, system_prompt, text_block_id
+            ):
+                yield response
+
+        # Save the session with accumulated messages
+        await upsert_chat_session(session)
+        logger.debug(
+            f"[SDK] Session {session_id} saved with {len(session.messages)} messages"
+        )
+        # Always yield StreamFinish to signal completion to the caller
+        # The adapter yields StreamFinish for the SSE stream, but we need to
+        # yield it here so the background task in routes.py knows to call mark_task_completed
+        yield StreamFinish()
+
+    except Exception as e:
+        logger.error(f"[SDK] Error: {e}", exc_info=True)
+        # Save session even on error to preserve any partial response
+        try:
+            await upsert_chat_session(session)
+        except Exception as save_err:
+            logger.error(f"[SDK] Failed to save session on error: {save_err}")
+        # Sanitize error message to avoid exposing internal details
+        yield StreamError(
+            errorText="An error occurred. Please try again.",
+            code="sdk_error",
+        )
+        yield StreamFinish()
+
+
+async def _update_title_async(
+    session_id: str, message: str, user_id: str | None = None
+) -> None:
+    """Background task to update session title."""
+    try:
+        title = await _generate_session_title(
+            message, user_id=user_id, session_id=session_id
+        )
+        if title:
+            await update_session_title(session_id, title)
+            logger.debug(f"[SDK] Generated title for {session_id}: {title}")
+    except Exception as e:
+        logger.warning(f"[SDK] Failed to update session title: {e}")

autogpt_platform/backend/backend/api/features/chat/sdk/tool_adapter.py

@@ -0,0 +1,213 @@
+"""Tool adapter for wrapping existing CoPilot tools as Claude Agent SDK MCP tools.
+
+This module provides the adapter layer that converts existing BaseTool implementations
+into in-process MCP tools that can be used with the Claude Agent SDK.
+"""
+
+import json
+import logging
+from contextvars import ContextVar
+from typing import Any
+
+from backend.api.features.chat.model import ChatSession
+from backend.api.features.chat.tools import TOOL_REGISTRY
+from backend.api.features.chat.tools.base import BaseTool
+
+logger = logging.getLogger(__name__)
+
+# Context variables to pass user/session info to tool execution
+_current_user_id: ContextVar[str | None] = ContextVar("current_user_id", default=None)
+_current_session: ContextVar[ChatSession | None] = ContextVar(
+    "current_session", default=None
+)
+_current_tool_call_id: ContextVar[str | None] = ContextVar(
+    "current_tool_call_id", default=None
+)
+
+
+def set_execution_context(
+    user_id: str | None,
+    session: ChatSession,
+    tool_call_id: str | None = None,
+) -> None:
+    """Set the execution context for tool calls.
+
+    This must be called before streaming begins to ensure tools have access
+    to user_id and session information.
+    """
+    _current_user_id.set(user_id)
+    _current_session.set(session)
+    _current_tool_call_id.set(tool_call_id)
+
+
+def get_execution_context() -> tuple[str | None, ChatSession | None, str | None]:
+    """Get the current execution context."""
+    return (
+        _current_user_id.get(),
+        _current_session.get(),
+        _current_tool_call_id.get(),
+    )
+
+
+def create_tool_handler(base_tool: BaseTool):
+    """Create an async handler function for a BaseTool.
+
+    This wraps the existing BaseTool._execute method to be compatible
+    with the Claude Agent SDK MCP tool format.
+    """
+
+    async def tool_handler(args: dict[str, Any]) -> dict[str, Any]:
+        """Execute the wrapped tool and return MCP-formatted response."""
+        user_id, session, tool_call_id = get_execution_context()
+
+        if session is None:
+            return {
+                "content": [
+                    {
+                        "type": "text",
+                        "text": json.dumps(
+                            {
+                                "error": "No session context available",
+                                "type": "error",
+                            }
+                        ),
+                    }
+                ],
+                "isError": True,
+            }
+
+        try:
+            # Call the existing tool's execute method
+            result = await base_tool.execute(
+                user_id=user_id,
+                session=session,
+                tool_call_id=tool_call_id or "sdk-call",
+                **args,
+            )
+
+            # The result is a StreamToolOutputAvailable, extract the output
+            return {
+                "content": [
+                    {
+                        "type": "text",
+                        "text": (
+                            result.output
+                            if isinstance(result.output, str)
+                            else json.dumps(result.output)
+                        ),
+                    }
+                ],
+                "isError": not result.success,
+            }
+
+        except Exception as e:
+            logger.error(f"Error executing tool {base_tool.name}: {e}", exc_info=True)
+            return {
+                "content": [
+                    {
+                        "type": "text",
+                        "text": json.dumps(
+                            {
+                                "error": str(e),
+                                "type": "error",
+                                "message": f"Failed to execute {base_tool.name}",
+                            }
+                        ),
+                    }
+                ],
+                "isError": True,
+            }
+
+    return tool_handler
+
+
+def get_tool_definitions() -> list[dict[str, Any]]:
+    """Get all tool definitions in MCP format.
+
+    Returns a list of tool definitions that can be used with
+    create_sdk_mcp_server or as raw tool definitions.
+    """
+    tool_definitions = []
+
+    for tool_name, base_tool in TOOL_REGISTRY.items():
+        tool_def = {
+            "name": tool_name,
+            "description": base_tool.description,
+            "inputSchema": {
+                "type": "object",
+                "properties": base_tool.parameters.get("properties", {}),
+                "required": base_tool.parameters.get("required", []),
+            },
+        }
+        tool_definitions.append(tool_def)
+
+    return tool_definitions
+
+
+def get_tool_handlers() -> dict[str, Any]:
+    """Get all tool handlers mapped by name.
+
+    Returns a dictionary mapping tool names to their handler functions.
+    """
+    handlers = {}
+
+    for tool_name, base_tool in TOOL_REGISTRY.items():
+        handlers[tool_name] = create_tool_handler(base_tool)
+
+    return handlers
+
+
+# Create the MCP server configuration
+def create_copilot_mcp_server():
+    """Create an in-process MCP server configuration for CoPilot tools.
+
+    This can be passed to ClaudeAgentOptions.mcp_servers.
+
+    Note: The actual SDK MCP server creation depends on the claude-agent-sdk
+    package being available. This function returns the configuration that
+    can be used with the SDK.
+    """
+    try:
+        from claude_agent_sdk import create_sdk_mcp_server, tool
+
+        # Create decorated tool functions
+        sdk_tools = []
+
+        for tool_name, base_tool in TOOL_REGISTRY.items():
+            # Get the handler
+            handler = create_tool_handler(base_tool)
+
+            # Create the decorated tool
+            # The @tool decorator expects (name, description, schema)
+            decorated = tool(
+                tool_name,
+                base_tool.description,
+                base_tool.parameters.get("properties", {}),
+            )(handler)
+
+            sdk_tools.append(decorated)
+
+        # Create the MCP server
+        server = create_sdk_mcp_server(
+            name="copilot",
+            version="1.0.0",
+            tools=sdk_tools,
+        )
+
+        return server
+
+    except ImportError:
+        logger.warning(
+            "claude-agent-sdk not available, returning tool definitions only"
+        )
+        return {
+            "tools": get_tool_definitions(),
+            "handlers": get_tool_handlers(),
+        }
+
+
+# List of tool names for allowed_tools configuration
+COPILOT_TOOL_NAMES = [f"mcp__copilot__{name}" for name in TOOL_REGISTRY.keys()]
+
+# Also export the raw tool names for flexibility
+RAW_TOOL_NAMES = list(TOOL_REGISTRY.keys())

autogpt_platform/backend/backend/api/features/chat/stream_registry.py

@@ -0,0 +1,708 @@
+"""Stream registry for managing reconnectable SSE streams.
+
+This module provides a registry for tracking active streaming tasks and their
+messages. It uses Redis for all state management (no in-memory state), making
+pods stateless and horizontally scalable.
+
+Architecture:
+- Redis Stream: Persists all messages for replay and real-time delivery
+- Redis Hash: Task metadata (status, session_id, etc.)
+
+Subscribers:
+1. Replay missed messages from Redis Stream (XREAD)
+2. Listen for live updates via blocking XREAD
+3. No in-memory state required on the subscribing pod
+"""
+
+import asyncio
+import logging
+from dataclasses import dataclass, field
+from datetime import datetime, timezone
+from typing import Any, Literal
+
+import orjson
+
+from backend.data.redis_client import get_redis_async
+
+from .config import ChatConfig
+from .response_model import StreamBaseResponse, StreamError, StreamFinish
+
+logger = logging.getLogger(__name__)
+config = ChatConfig()
+
+# Track background tasks for this pod (just the asyncio.Task reference, not subscribers)
+_local_tasks: dict[str, asyncio.Task] = {}
+
+# Track listener tasks per subscriber queue for cleanup
+# Maps queue id() to (task_id, asyncio.Task) for proper cleanup on unsubscribe
+_listener_tasks: dict[int, tuple[str, asyncio.Task]] = {}
+
+# Timeout for putting chunks into subscriber queues (seconds)
+# If the queue is full and doesn't drain within this time, send an overflow error
+QUEUE_PUT_TIMEOUT = 5.0
+
+# Lua script for atomic compare-and-swap status update (idempotent completion)
+# Returns 1 if status was updated, 0 if already completed/failed
+COMPLETE_TASK_SCRIPT = """
+local current = redis.call("HGET", KEYS[1], "status")
+if current == "running" then
+    redis.call("HSET", KEYS[1], "status", ARGV[1])
+    return 1
+end
+return 0
+"""
+
+
+@dataclass
+class ActiveTask:
+    """Represents an active streaming task (metadata only, no in-memory queues)."""
+
+    task_id: str
+    session_id: str
+    user_id: str | None
+    tool_call_id: str
+    tool_name: str
+    operation_id: str
+    status: Literal["running", "completed", "failed"] = "running"
+    created_at: datetime = field(default_factory=lambda: datetime.now(timezone.utc))
+    asyncio_task: asyncio.Task | None = None
+
+
+def _get_task_meta_key(task_id: str) -> str:
+    """Get Redis key for task metadata."""
+    return f"{config.task_meta_prefix}{task_id}"
+
+
+def _get_task_stream_key(task_id: str) -> str:
+    """Get Redis key for task message stream."""
+    return f"{config.task_stream_prefix}{task_id}"
+
+
+def _get_operation_mapping_key(operation_id: str) -> str:
+    """Get Redis key for operation_id to task_id mapping."""
+    return f"{config.task_op_prefix}{operation_id}"
+
+
+async def create_task(
+    task_id: str,
+    session_id: str,
+    user_id: str | None,
+    tool_call_id: str,
+    tool_name: str,
+    operation_id: str,
+) -> ActiveTask:
+    """Create a new streaming task in Redis.
+
+    Args:
+        task_id: Unique identifier for the task
+        session_id: Chat session ID
+        user_id: User ID (may be None for anonymous)
+        tool_call_id: Tool call ID from the LLM
+        tool_name: Name of the tool being executed
+        operation_id: Operation ID for webhook callbacks
+
+    Returns:
+        The created ActiveTask instance (metadata only)
+    """
+    task = ActiveTask(
+        task_id=task_id,
+        session_id=session_id,
+        user_id=user_id,
+        tool_call_id=tool_call_id,
+        tool_name=tool_name,
+        operation_id=operation_id,
+    )
+
+    # Store metadata in Redis
+    redis = await get_redis_async()
+    meta_key = _get_task_meta_key(task_id)
+    op_key = _get_operation_mapping_key(operation_id)
+
+    await redis.hset(  # type: ignore[misc]
+        meta_key,
+        mapping={
+            "task_id": task_id,
+            "session_id": session_id,
+            "user_id": user_id or "",
+            "tool_call_id": tool_call_id,
+            "tool_name": tool_name,
+            "operation_id": operation_id,
+            "status": task.status,
+            "created_at": task.created_at.isoformat(),
+        },
+    )
+    await redis.expire(meta_key, config.stream_ttl)
+
+    # Create operation_id -> task_id mapping for webhook lookups
+    await redis.set(op_key, task_id, ex=config.stream_ttl)
+
+    logger.debug(f"Created task {task_id} for session {session_id}")
+
+    return task
+
+
+async def publish_chunk(
+    task_id: str,
+    chunk: StreamBaseResponse,
+) -> str:
+    """Publish a chunk to Redis Stream.
+
+    All delivery is via Redis Streams - no in-memory state.
+
+    Args:
+        task_id: Task ID to publish to
+        chunk: The stream response chunk to publish
+
+    Returns:
+        The Redis Stream message ID
+    """
+    chunk_json = chunk.model_dump_json()
+    message_id = "0-0"
+
+    try:
+        redis = await get_redis_async()
+        stream_key = _get_task_stream_key(task_id)
+
+        # Write to Redis Stream for persistence and real-time delivery
+        raw_id = await redis.xadd(
+            stream_key,
+            {"data": chunk_json},
+            maxlen=config.stream_max_length,
+        )
+        message_id = raw_id if isinstance(raw_id, str) else raw_id.decode()
+
+        # Set TTL on stream to match task metadata TTL
+        await redis.expire(stream_key, config.stream_ttl)
+    except Exception as e:
+        logger.error(
+            f"Failed to publish chunk for task {task_id}: {e}",
+            exc_info=True,
+        )
+
+    return message_id
+
+
+async def subscribe_to_task(
+    task_id: str,
+    user_id: str | None,
+    last_message_id: str = "0-0",
+) -> asyncio.Queue[StreamBaseResponse] | None:
+    """Subscribe to a task's stream with replay of missed messages.
+
+    This is fully stateless - uses Redis Stream for replay and pub/sub for live updates.
+
+    Args:
+        task_id: Task ID to subscribe to
+        user_id: User ID for ownership validation
+        last_message_id: Last Redis Stream message ID received ("0-0" for full replay)
+
+    Returns:
+        An asyncio Queue that will receive stream chunks, or None if task not found
+        or user doesn't have access
+    """
+    redis = await get_redis_async()
+    meta_key = _get_task_meta_key(task_id)
+    meta: dict[Any, Any] = await redis.hgetall(meta_key)  # type: ignore[misc]
+
+    if not meta:
+        logger.debug(f"Task {task_id} not found in Redis")
+        return None
+
+    # Note: Redis client uses decode_responses=True, so keys are strings
+    task_status = meta.get("status", "")
+    task_user_id = meta.get("user_id", "") or None
+
+    # Validate ownership - if task has an owner, requester must match
+    if task_user_id:
+        if user_id != task_user_id:
+            logger.warning(
+                f"User {user_id} denied access to task {task_id} "
+                f"owned by {task_user_id}"
+            )
+            return None
+
+    subscriber_queue: asyncio.Queue[StreamBaseResponse] = asyncio.Queue()
+    stream_key = _get_task_stream_key(task_id)
+
+    # Step 1: Replay messages from Redis Stream
+    messages = await redis.xread({stream_key: last_message_id}, block=0, count=1000)
+
+    replayed_count = 0
+    replay_last_id = last_message_id
+    if messages:
+        for _stream_name, stream_messages in messages:
+            for msg_id, msg_data in stream_messages:
+                replay_last_id = msg_id if isinstance(msg_id, str) else msg_id.decode()
+                # Note: Redis client uses decode_responses=True, so keys are strings
+                if "data" in msg_data:
+                    try:
+                        chunk_data = orjson.loads(msg_data["data"])
+                        chunk = _reconstruct_chunk(chunk_data)
+                        if chunk:
+                            await subscriber_queue.put(chunk)
+                            replayed_count += 1
+                    except Exception as e:
+                        logger.warning(f"Failed to replay message: {e}")
+
+    logger.debug(f"Task {task_id}: replayed {replayed_count} messages")
+
+    # Step 2: If task is still running, start stream listener for live updates
+    if task_status == "running":
+        listener_task = asyncio.create_task(
+            _stream_listener(task_id, subscriber_queue, replay_last_id)
+        )
+        # Track listener task for cleanup on unsubscribe
+        _listener_tasks[id(subscriber_queue)] = (task_id, listener_task)
+    else:
+        # Task is completed/failed - add finish marker
+        await subscriber_queue.put(StreamFinish())
+
+    return subscriber_queue
+
+
+async def _stream_listener(
+    task_id: str,
+    subscriber_queue: asyncio.Queue[StreamBaseResponse],
+    last_replayed_id: str,
+) -> None:
+    """Listen to Redis Stream for new messages using blocking XREAD.
+
+    This approach avoids the duplicate message issue that can occur with pub/sub
+    when messages are published during the gap between replay and subscription.
+
+    Args:
+        task_id: Task ID to listen for
+        subscriber_queue: Queue to deliver messages to
+        last_replayed_id: Last message ID from replay (continue from here)
+    """
+    queue_id = id(subscriber_queue)
+    # Track the last successfully delivered message ID for recovery hints
+    last_delivered_id = last_replayed_id
+
+    try:
+        redis = await get_redis_async()
+        stream_key = _get_task_stream_key(task_id)
+        current_id = last_replayed_id
+
+        while True:
+            # Block for up to 30 seconds waiting for new messages
+            # This allows periodic checking if task is still running
+            messages = await redis.xread(
+                {stream_key: current_id}, block=30000, count=100
+            )
+
+            if not messages:
+                # Timeout - check if task is still running
+                meta_key = _get_task_meta_key(task_id)
+                status = await redis.hget(meta_key, "status")  # type: ignore[misc]
+                if status and status != "running":
+                    try:
+                        await asyncio.wait_for(
+                            subscriber_queue.put(StreamFinish()),
+                            timeout=QUEUE_PUT_TIMEOUT,
+                        )
+                    except asyncio.TimeoutError:
+                        logger.warning(
+                            f"Timeout delivering finish event for task {task_id}"
+                        )
+                    break
+                continue
+
+            for _stream_name, stream_messages in messages:
+                for msg_id, msg_data in stream_messages:
+                    current_id = msg_id if isinstance(msg_id, str) else msg_id.decode()
+
+                    if "data" not in msg_data:
+                        continue
+
+                    try:
+                        chunk_data = orjson.loads(msg_data["data"])
+                        chunk = _reconstruct_chunk(chunk_data)
+                        if chunk:
+                            try:
+                                await asyncio.wait_for(
+                                    subscriber_queue.put(chunk),
+                                    timeout=QUEUE_PUT_TIMEOUT,
+                                )
+                                # Update last delivered ID on successful delivery
+                                last_delivered_id = current_id
+                            except asyncio.TimeoutError:
+                                logger.warning(
+                                    f"Subscriber queue full for task {task_id}, "
+                                    f"message delivery timed out after {QUEUE_PUT_TIMEOUT}s"
+                                )
+                                # Send overflow error with recovery info
+                                try:
+                                    overflow_error = StreamError(
+                                        errorText="Message delivery timeout - some messages may have been missed",
+                                        code="QUEUE_OVERFLOW",
+                                        details={
+                                            "last_delivered_id": last_delivered_id,
+                                            "recovery_hint": f"Reconnect with last_message_id={last_delivered_id}",
+                                        },
+                                    )
+                                    subscriber_queue.put_nowait(overflow_error)
+                                except asyncio.QueueFull:
+                                    # Queue is completely stuck, nothing more we can do
+                                    logger.error(
+                                        f"Cannot deliver overflow error for task {task_id}, "
+                                        "queue completely blocked"
+                                    )
+
+                            # Stop listening on finish
+                            if isinstance(chunk, StreamFinish):
+                                return
+                    except Exception as e:
+                        logger.warning(f"Error processing stream message: {e}")
+
+    except asyncio.CancelledError:
+        logger.debug(f"Stream listener cancelled for task {task_id}")
+        raise  # Re-raise to propagate cancellation
+    except Exception as e:
+        logger.error(f"Stream listener error for task {task_id}: {e}")
+        # On error, send finish to unblock subscriber
+        try:
+            await asyncio.wait_for(
+                subscriber_queue.put(StreamFinish()),
+                timeout=QUEUE_PUT_TIMEOUT,
+            )
+        except (asyncio.TimeoutError, asyncio.QueueFull):
+            logger.warning(
+                f"Could not deliver finish event for task {task_id} after error"
+            )
+    finally:
+        # Clean up listener task mapping on exit
+        _listener_tasks.pop(queue_id, None)
+
+
+async def mark_task_completed(
+    task_id: str,
+    status: Literal["completed", "failed"] = "completed",
+) -> bool:
+    """Mark a task as completed and publish finish event.
+
+    This is idempotent - calling multiple times with the same task_id is safe.
+    Uses atomic compare-and-swap via Lua script to prevent race conditions.
+    Status is updated first (source of truth), then finish event is published (best-effort).
+
+    Args:
+        task_id: Task ID to mark as completed
+        status: Final status ("completed" or "failed")
+
+    Returns:
+        True if task was newly marked completed, False if already completed/failed
+    """
+    redis = await get_redis_async()
+    meta_key = _get_task_meta_key(task_id)
+
+    # Atomic compare-and-swap: only update if status is "running"
+    # This prevents race conditions when multiple callers try to complete simultaneously
+    result = await redis.eval(COMPLETE_TASK_SCRIPT, 1, meta_key, status)  # type: ignore[misc]
+
+    if result == 0:
+        logger.debug(f"Task {task_id} already completed/failed, skipping")
+        return False
+
+    # THEN publish finish event (best-effort - listeners can detect via status polling)
+    try:
+        await publish_chunk(task_id, StreamFinish())
+    except Exception as e:
+        logger.error(
+            f"Failed to publish finish event for task {task_id}: {e}. "
+            "Listeners will detect completion via status polling."
+        )
+
+    # Clean up local task reference if exists
+    _local_tasks.pop(task_id, None)
+    return True
+
+
+async def find_task_by_operation_id(operation_id: str) -> ActiveTask | None:
+    """Find a task by its operation ID.
+
+    Used by webhook callbacks to locate the task to update.
+
+    Args:
+        operation_id: Operation ID to search for
+
+    Returns:
+        ActiveTask if found, None otherwise
+    """
+    redis = await get_redis_async()
+    op_key = _get_operation_mapping_key(operation_id)
+    task_id = await redis.get(op_key)
+
+    if not task_id:
+        return None
+
+    task_id_str = task_id.decode() if isinstance(task_id, bytes) else task_id
+    return await get_task(task_id_str)
+
+
+async def get_task(task_id: str) -> ActiveTask | None:
+    """Get a task by its ID from Redis.
+
+    Args:
+        task_id: Task ID to look up
+
+    Returns:
+        ActiveTask if found, None otherwise
+    """
+    redis = await get_redis_async()
+    meta_key = _get_task_meta_key(task_id)
+    meta: dict[Any, Any] = await redis.hgetall(meta_key)  # type: ignore[misc]
+
+    if not meta:
+        return None
+
+    # Note: Redis client uses decode_responses=True, so keys/values are strings
+    return ActiveTask(
+        task_id=meta.get("task_id", ""),
+        session_id=meta.get("session_id", ""),
+        user_id=meta.get("user_id", "") or None,
+        tool_call_id=meta.get("tool_call_id", ""),
+        tool_name=meta.get("tool_name", ""),
+        operation_id=meta.get("operation_id", ""),
+        status=meta.get("status", "running"),  # type: ignore[arg-type]
+    )
+
+
+async def get_task_with_expiry_info(
+    task_id: str,
+) -> tuple[ActiveTask | None, str | None]:
+    """Get a task by its ID with expiration detection.
+
+    Returns (task, error_code) where error_code is:
+    - None if task found
+    - "TASK_EXPIRED" if stream exists but metadata is gone (TTL expired)
+    - "TASK_NOT_FOUND" if neither exists
+
+    Args:
+        task_id: Task ID to look up
+
+    Returns:
+        Tuple of (ActiveTask or None, error_code or None)
+    """
+    redis = await get_redis_async()
+    meta_key = _get_task_meta_key(task_id)
+    stream_key = _get_task_stream_key(task_id)
+
+    meta: dict[Any, Any] = await redis.hgetall(meta_key)  # type: ignore[misc]
+
+    if not meta:
+        # Check if stream still has data (metadata expired but stream hasn't)
+        stream_len = await redis.xlen(stream_key)
+        if stream_len > 0:
+            return None, "TASK_EXPIRED"
+        return None, "TASK_NOT_FOUND"
+
+    # Note: Redis client uses decode_responses=True, so keys/values are strings
+    return (
+        ActiveTask(
+            task_id=meta.get("task_id", ""),
+            session_id=meta.get("session_id", ""),
+            user_id=meta.get("user_id", "") or None,
+            tool_call_id=meta.get("tool_call_id", ""),
+            tool_name=meta.get("tool_name", ""),
+            operation_id=meta.get("operation_id", ""),
+            status=meta.get("status", "running"),  # type: ignore[arg-type]
+        ),
+        None,
+    )
+
+
+async def get_active_task_for_session(
+    session_id: str,
+    user_id: str | None = None,
+) -> tuple[ActiveTask | None, str]:
+    """Get the active (running) task for a session, if any.
+
+    Scans Redis for tasks matching the session_id with status="running".
+
+    Args:
+        session_id: Session ID to look up
+        user_id: User ID for ownership validation (optional)
+
+    Returns:
+        Tuple of (ActiveTask if found and running, last_message_id from Redis Stream)
+    """
+
+    redis = await get_redis_async()
+
+    # Scan Redis for task metadata keys
+    cursor = 0
+    tasks_checked = 0
+
+    while True:
+        cursor, keys = await redis.scan(
+            cursor, match=f"{config.task_meta_prefix}*", count=100
+        )
+
+        for key in keys:
+            tasks_checked += 1
+            meta: dict[Any, Any] = await redis.hgetall(key)  # type: ignore[misc]
+            if not meta:
+                continue
+
+            # Note: Redis client uses decode_responses=True, so keys/values are strings
+            task_session_id = meta.get("session_id", "")
+            task_status = meta.get("status", "")
+            task_user_id = meta.get("user_id", "") or None
+            task_id = meta.get("task_id", "")
+
+            if task_session_id == session_id and task_status == "running":
+                # Validate ownership - if task has an owner, requester must match
+                if task_user_id and user_id != task_user_id:
+                    continue
+
+                logger.info(
+                    f"[TASK_LOOKUP] Found running task {task_id[:8]}... for session {session_id[:8]}..."
+                )
+
+                # Get the last message ID from Redis Stream
+                stream_key = _get_task_stream_key(task_id)
+                last_id = "0-0"
+                try:
+                    messages = await redis.xrevrange(stream_key, count=1)
+                    if messages:
+                        msg_id = messages[0][0]
+                        last_id = msg_id if isinstance(msg_id, str) else msg_id.decode()
+                except Exception as e:
+                    logger.warning(f"Failed to get last message ID: {e}")
+
+                return (
+                    ActiveTask(
+                        task_id=task_id,
+                        session_id=task_session_id,
+                        user_id=task_user_id,
+                        tool_call_id=meta.get("tool_call_id", ""),
+                        tool_name=meta.get("tool_name", ""),
+                        operation_id=meta.get("operation_id", ""),
+                        status="running",
+                    ),
+                    last_id,
+                )
+
+        if cursor == 0:
+            break
+
+    return None, "0-0"
+
+
+def _reconstruct_chunk(chunk_data: dict) -> StreamBaseResponse | None:
+    """Reconstruct a StreamBaseResponse from JSON data.
+
+    Args:
+        chunk_data: Parsed JSON data from Redis
+
+    Returns:
+        Reconstructed response object, or None if unknown type
+    """
+    from .response_model import (
+        ResponseType,
+        StreamError,
+        StreamFinish,
+        StreamHeartbeat,
+        StreamStart,
+        StreamTextDelta,
+        StreamTextEnd,
+        StreamTextStart,
+        StreamToolInputAvailable,
+        StreamToolInputStart,
+        StreamToolOutputAvailable,
+        StreamUsage,
+    )
+
+    # Map response types to their corresponding classes
+    type_to_class: dict[str, type[StreamBaseResponse]] = {
+        ResponseType.START.value: StreamStart,
+        ResponseType.FINISH.value: StreamFinish,
+        ResponseType.TEXT_START.value: StreamTextStart,
+        ResponseType.TEXT_DELTA.value: StreamTextDelta,
+        ResponseType.TEXT_END.value: StreamTextEnd,
+        ResponseType.TOOL_INPUT_START.value: StreamToolInputStart,
+        ResponseType.TOOL_INPUT_AVAILABLE.value: StreamToolInputAvailable,
+        ResponseType.TOOL_OUTPUT_AVAILABLE.value: StreamToolOutputAvailable,
+        ResponseType.ERROR.value: StreamError,
+        ResponseType.USAGE.value: StreamUsage,
+        ResponseType.HEARTBEAT.value: StreamHeartbeat,
+    }
+
+    chunk_type = chunk_data.get("type")
+    chunk_class = type_to_class.get(chunk_type)  # type: ignore[arg-type]
+
+    if chunk_class is None:
+        logger.warning(f"Unknown chunk type: {chunk_type}")
+        return None
+
+    try:
+        return chunk_class(**chunk_data)
+    except Exception as e:
+        logger.warning(f"Failed to reconstruct chunk of type {chunk_type}: {e}")
+        return None
+
+
+async def set_task_asyncio_task(task_id: str, asyncio_task: asyncio.Task) -> None:
+    """Track the asyncio.Task for a task (local reference only).
+
+    This is just for cleanup purposes - the task state is in Redis.
+
+    Args:
+        task_id: Task ID
+        asyncio_task: The asyncio Task to track
+    """
+    _local_tasks[task_id] = asyncio_task
+
+
+async def unsubscribe_from_task(
+    task_id: str,
+    subscriber_queue: asyncio.Queue[StreamBaseResponse],
+) -> None:
+    """Clean up when a subscriber disconnects.
+
+    Cancels the XREAD-based listener task associated with this subscriber queue
+    to prevent resource leaks.
+
+    Args:
+        task_id: Task ID
+        subscriber_queue: The subscriber's queue used to look up the listener task
+    """
+    queue_id = id(subscriber_queue)
+    listener_entry = _listener_tasks.pop(queue_id, None)
+
+    if listener_entry is None:
+        logger.debug(
+            f"No listener task found for task {task_id} queue {queue_id} "
+            "(may have already completed)"
+        )
+        return
+
+    stored_task_id, listener_task = listener_entry
+
+    if stored_task_id != task_id:
+        logger.warning(
+            f"Task ID mismatch in unsubscribe: expected {task_id}, "
+            f"found {stored_task_id}"
+        )
+
+    if listener_task.done():
+        logger.debug(f"Listener task for task {task_id} already completed")
+        return
+
+    # Cancel the listener task
+    listener_task.cancel()
+
+    try:
+        # Wait for the task to be cancelled with a timeout
+        await asyncio.wait_for(listener_task, timeout=5.0)
+    except asyncio.CancelledError:
+        # Expected - the task was successfully cancelled
+        pass
+    except asyncio.TimeoutError:
+        logger.warning(
+            f"Timeout waiting for listener task cancellation for task {task_id}"
+        )
+    except Exception as e:
+        logger.error(f"Error during listener task cancellation for task {task_id}: {e}")
+
+    logger.debug(f"Successfully unsubscribed from task {task_id}")

autogpt_platform/backend/backend/api/features/chat/tools/__init__.py

@@ -10,6 +10,7 @@ from .add_understanding import AddUnderstandingTool
 from .agent_output import AgentOutputTool
 from .base import BaseTool
 from .create_agent import CreateAgentTool
+from .customize_agent import CustomizeAgentTool
 from .edit_agent import EditAgentTool
 from .find_agent import FindAgentTool
 from .find_block import FindBlockTool
@@ -34,6 +35,7 @@ logger = logging.getLogger(__name__)
 TOOL_REGISTRY: dict[str, BaseTool] = {
     "add_understanding": AddUnderstandingTool(),
     "create_agent": CreateAgentTool(),
+    "customize_agent": CustomizeAgentTool(),
     "edit_agent": EditAgentTool(),
     "find_agent": FindAgentTool(),
     "find_block": FindBlockTool(),

autogpt_platform/backend/backend/api/features/chat/tools/agent_generator/__init__.py

@@ -8,6 +8,7 @@ from .core import (
     DecompositionStep,
     LibraryAgentSummary,
     MarketplaceAgentSummary,
+    customize_template,
     decompose_goal,
     enrich_library_agents_from_steps,
     extract_search_terms_from_steps,
@@ -19,6 +20,7 @@ from .core import (
     get_library_agent_by_graph_id,
     get_library_agent_by_id,
     get_library_agents_for_generation,
+    graph_to_json,
     json_to_graph,
     save_agent_to_library,
     search_marketplace_agents_for_generation,
@@ -36,6 +38,7 @@ __all__ = [
     "LibraryAgentSummary",
     "MarketplaceAgentSummary",
     "check_external_service_health",
+    "customize_template",
     "decompose_goal",
     "enrich_library_agents_from_steps",
     "extract_search_terms_from_steps",
@@ -48,6 +51,7 @@ __all__ = [
     "get_library_agent_by_id",
     "get_library_agents_for_generation",
     "get_user_message_for_error",
+    "graph_to_json",
     "is_external_service_configured",
     "json_to_graph",
     "save_agent_to_library",

autogpt_platform/backend/backend/api/features/chat/tools/agent_generator/core.py

@@ -7,17 +7,11 @@ from typing import Any, NotRequired, TypedDict
 
 from backend.api.features.library import db as library_db
 from backend.api.features.store import db as store_db
-from backend.data.graph import (
+from backend.data.graph import Graph, Link, Node, get_graph, get_store_listed_graphs
-    Graph,
-    Link,
-    Node,
-    create_graph,
-    get_graph,
-    get_graph_all_versions,
-)
 from backend.util.exceptions import DatabaseError, NotFoundError
 
 from .service import (
+    customize_template_external,
     decompose_goal_external,
     generate_agent_external,
     generate_agent_patch_external,
@@ -26,8 +20,6 @@ from .service import (
 
 logger = logging.getLogger(__name__)
 
-AGENT_EXECUTOR_BLOCK_ID = "e189baac-8c20-45a1-94a7-55177ea42565"
-
 
 class ExecutionSummary(TypedDict):
     """Summary of a single execution for quality assessment."""
@@ -266,18 +258,18 @@ async def get_library_agents_for_generation(
 async def search_marketplace_agents_for_generation(
     search_query: str,
     max_results: int = 10,
-) -> list[MarketplaceAgentSummary]:
+) -> list[LibraryAgentSummary]:
     """Search marketplace agents formatted for Agent Generator.
 
-    Note: This returns basic agent info. Full input/output schemas would require
+    Fetches marketplace agents and their full schemas so they can be used
-    additional graph fetches and is a potential future enhancement.
+    as sub-agents in generated workflows.
 
     Args:
         search_query: Search term to find relevant public agents
         max_results: Maximum number of agents to return (default 10)
 
     Returns:
-        List of MarketplaceAgentSummary (without detailed schemas for now)
+        List of LibraryAgentSummary with full input/output schemas
     """
     try:
         response = await store_db.get_store_agents(
@@ -286,17 +278,31 @@ async def search_marketplace_agents_for_generation(
             page_size=max_results,
         )
 
-        results: list[MarketplaceAgentSummary] = []
+        agents_with_graphs = [
-        for agent in response.agents:
+            agent for agent in response.agents if agent.agent_graph_id
-            results.append(
+        ]
-                MarketplaceAgentSummary(
+
-                    name=agent.agent_name,
+        if not agents_with_graphs:
-                    description=agent.description,
+            return []
-                    sub_heading=agent.sub_heading,
+
-                    creator=agent.creator,
+        graph_ids = [agent.agent_graph_id for agent in agents_with_graphs]
-                    is_marketplace_agent=True,
+        graphs = await get_store_listed_graphs(*graph_ids)
+
+        results: list[LibraryAgentSummary] = []
+        for agent in agents_with_graphs:
+            graph_id = agent.agent_graph_id
+            if graph_id and graph_id in graphs:
+                graph = graphs[graph_id]
+                results.append(
+                    LibraryAgentSummary(
+                        graph_id=graph.id,
+                        graph_version=graph.version,
+                        name=agent.agent_name,
+                        description=agent.description,
+                        input_schema=graph.input_schema,
+                        output_schema=graph.output_schema,
+                    )
                 )
-            )
         return results
     except Exception as e:
         logger.warning(f"Failed to search marketplace agents: {e}")
@@ -327,8 +333,7 @@ async def get_all_relevant_agents_for_generation(
         max_marketplace_results: Max marketplace agents to return (default 10)
 
     Returns:
-        List of AgentSummary, library agents first (with full schemas),
+        List of AgentSummary with full schemas (both library and marketplace agents)
-        then marketplace agents (basic info only)
     """
     agents: list[AgentSummary] = []
     seen_graph_ids: set[str] = set()
@@ -365,16 +370,11 @@ async def get_all_relevant_agents_for_generation(
             search_query=search_query,
             max_results=max_marketplace_results,
         )
-        library_names: set[str] = set()
-        for a in agents:
-            name = a.get("name")
-            if name and isinstance(name, str):
-                library_names.add(name.lower())
         for agent in marketplace_agents:
-            agent_name = agent.get("name")
+            graph_id = agent.get("graph_id")
-            if agent_name and isinstance(agent_name, str):
+            if graph_id and graph_id not in seen_graph_ids:
-                if agent_name.lower() not in library_names:
+                agents.append(agent)
-                    agents.append(agent)
+                seen_graph_ids.add(graph_id)
 
     return agents
 
@@ -540,15 +540,21 @@ async def decompose_goal(
 async def generate_agent(
     instructions: DecompositionResult | dict[str, Any],
     library_agents: list[AgentSummary] | list[dict[str, Any]] | None = None,
+    operation_id: str | None = None,
+    task_id: str | None = None,
 ) -> dict[str, Any] | None:
     """Generate agent JSON from instructions.
 
     Args:
         instructions: Structured instructions from decompose_goal
         library_agents: User's library agents available for sub-agent composition
+        operation_id: Operation ID for async processing (enables Redis Streams
+            completion notification)
+        task_id: Task ID for async processing (enables Redis Streams persistence
+            and SSE delivery)
 
     Returns:
-        Agent JSON dict, error dict {"type": "error", ...}, or None on error
+        Agent JSON dict, {"status": "accepted"} for async, error dict {"type": "error", ...}, or None on error
 
     Raises:
         AgentGeneratorNotConfiguredError: If the external service is not configured.
@@ -556,8 +562,13 @@ async def generate_agent(
     _check_service_configured()
     logger.info("Calling external Agent Generator service for generate_agent")
     result = await generate_agent_external(
-        dict(instructions), _to_dict_list(library_agents)
+        dict(instructions), _to_dict_list(library_agents), operation_id, task_id
     )
+
+    # Don't modify async response
+    if result and result.get("status") == "accepted":
+        return result
+
     if result:
         if isinstance(result, dict) and result.get("type") == "error":
             return result
@@ -648,45 +659,6 @@ def json_to_graph(agent_json: dict[str, Any]) -> Graph:
     )
 
 
-def _reassign_node_ids(graph: Graph) -> None:
-    """Reassign all node and link IDs to new UUIDs.
-
-    This is needed when creating a new version to avoid unique constraint violations.
-    """
-    id_map = {node.id: str(uuid.uuid4()) for node in graph.nodes}
-
-    for node in graph.nodes:
-        node.id = id_map[node.id]
-
-    for link in graph.links:
-        link.id = str(uuid.uuid4())
-        if link.source_id in id_map:
-            link.source_id = id_map[link.source_id]
-        if link.sink_id in id_map:
-            link.sink_id = id_map[link.sink_id]
-
-
-def _populate_agent_executor_user_ids(agent_json: dict[str, Any], user_id: str) -> None:
-    """Populate user_id in AgentExecutorBlock nodes.
-
-    The external agent generator creates AgentExecutorBlock nodes with empty user_id.
-    This function fills in the actual user_id so sub-agents run with correct permissions.
-
-    Args:
-        agent_json: Agent JSON dict (modified in place)
-        user_id: User ID to set
-    """
-    for node in agent_json.get("nodes", []):
-        if node.get("block_id") == AGENT_EXECUTOR_BLOCK_ID:
-            input_default = node.get("input_default") or {}
-            if not input_default.get("user_id"):
-                input_default["user_id"] = user_id
-                node["input_default"] = input_default
-                logger.debug(
-                    f"Set user_id for AgentExecutorBlock node {node.get('id')}"
-                )
-
-
 async def save_agent_to_library(
     agent_json: dict[str, Any], user_id: str, is_update: bool = False
 ) -> tuple[Graph, Any]:
@@ -700,63 +672,21 @@ async def save_agent_to_library(
     Returns:
         Tuple of (created Graph, LibraryAgent)
     """
-    # Populate user_id in AgentExecutorBlock nodes before conversion
-    _populate_agent_executor_user_ids(agent_json, user_id)
-
     graph = json_to_graph(agent_json)
-
     if is_update:
-        if graph.id:
+        return await library_db.update_graph_in_library(graph, user_id)
-            existing_versions = await get_graph_all_versions(graph.id, user_id)
+    return await library_db.create_graph_in_library(graph, user_id)
-            if existing_versions:
-                latest_version = max(v.version for v in existing_versions)
-                graph.version = latest_version + 1
-                _reassign_node_ids(graph)
-                logger.info(f"Updating agent {graph.id} to version {graph.version}")
-    else:
-        graph.id = str(uuid.uuid4())
-        graph.version = 1
-        _reassign_node_ids(graph)
-        logger.info(f"Creating new agent with ID {graph.id}")
-
-    created_graph = await create_graph(graph, user_id)
-
-    library_agents = await library_db.create_library_agent(
-        graph=created_graph,
-        user_id=user_id,
-        sensitive_action_safe_mode=True,
-        create_library_agents_for_sub_graphs=False,
-    )
-
-    return created_graph, library_agents[0]
 
 
-async def get_agent_as_json(
+def graph_to_json(graph: Graph) -> dict[str, Any]:
-    agent_id: str, user_id: str | None
+    """Convert a Graph object to JSON format for the agent generator.
-) -> dict[str, Any] | None:
-    """Fetch an agent and convert to JSON format for editing.
 
     Args:
-        agent_id: Graph ID or library agent ID
+        graph: Graph object to convert
-        user_id: User ID
 
     Returns:
-        Agent as JSON dict or None if not found
+        Agent as JSON dict
     """
-    graph = await get_graph(agent_id, version=None, user_id=user_id)
-
-    if not graph and user_id:
-        try:
-            library_agent = await library_db.get_library_agent(agent_id, user_id)
-            graph = await get_graph(
-                library_agent.graph_id, version=None, user_id=user_id
-            )
-        except NotFoundError:
-            pass
-
-    if not graph:
-        return None
-
     nodes = []
     for node in graph.nodes:
         nodes.append(
@@ -793,10 +723,41 @@ async def get_agent_as_json(
     }
 
 
+async def get_agent_as_json(
+    agent_id: str, user_id: str | None
+) -> dict[str, Any] | None:
+    """Fetch an agent and convert to JSON format for editing.
+
+    Args:
+        agent_id: Graph ID or library agent ID
+        user_id: User ID
+
+    Returns:
+        Agent as JSON dict or None if not found
+    """
+    graph = await get_graph(agent_id, version=None, user_id=user_id)
+
+    if not graph and user_id:
+        try:
+            library_agent = await library_db.get_library_agent(agent_id, user_id)
+            graph = await get_graph(
+                library_agent.graph_id, version=None, user_id=user_id
+            )
+        except NotFoundError:
+            pass
+
+    if not graph:
+        return None
+
+    return graph_to_json(graph)
+
+
 async def generate_agent_patch(
     update_request: str,
     current_agent: dict[str, Any],
     library_agents: list[AgentSummary] | None = None,
+    operation_id: str | None = None,
+    task_id: str | None = None,
 ) -> dict[str, Any] | None:
     """Update an existing agent using natural language.
 
@@ -809,10 +770,12 @@ async def generate_agent_patch(
         update_request: Natural language description of changes
         current_agent: Current agent JSON
         library_agents: User's library agents available for sub-agent composition
+        operation_id: Operation ID for async processing (enables Redis Streams callback)
+        task_id: Task ID for async processing (enables Redis Streams callback)
 
     Returns:
         Updated agent JSON, clarifying questions dict {"type": "clarifying_questions", ...},
-        error dict {"type": "error", ...}, or None on unexpected error
+        {"status": "accepted"} for async, error dict {"type": "error", ...}, or None on error
 
     Raises:
         AgentGeneratorNotConfiguredError: If the external service is not configured.
@@ -820,5 +783,43 @@ async def generate_agent_patch(
     _check_service_configured()
     logger.info("Calling external Agent Generator service for generate_agent_patch")
     return await generate_agent_patch_external(
-        update_request, current_agent, _to_dict_list(library_agents)
+        update_request,
+        current_agent,
+        _to_dict_list(library_agents),
+        operation_id,
+        task_id,
+    )
+
+
+async def customize_template(
+    template_agent: dict[str, Any],
+    modification_request: str,
+    context: str = "",
+) -> dict[str, Any] | None:
+    """Customize a template/marketplace agent using natural language.
+
+    This is used when users want to modify a template or marketplace agent
+    to fit their specific needs before adding it to their library.
+
+    The external Agent Generator service handles:
+    - Understanding the modification request
+    - Applying changes to the template
+    - Fixing and validating the result
+
+    Args:
+        template_agent: The template agent JSON to customize
+        modification_request: Natural language description of customizations
+        context: Additional context (e.g., answers to previous questions)
+
+    Returns:
+        Customized agent JSON, clarifying questions dict {"type": "clarifying_questions", ...},
+        error dict {"type": "error", ...}, or None on unexpected error
+
+    Raises:
+        AgentGeneratorNotConfiguredError: If the external service is not configured.
+    """
+    _check_service_configured()
+    logger.info("Calling external Agent Generator service for customize_template")
+    return await customize_template_external(
+        template_agent, modification_request, context
     )

autogpt_platform/backend/backend/api/features/chat/tools/agent_generator/service.py

@@ -139,11 +139,10 @@ async def decompose_goal_external(
     """
     client = _get_client()
 
-    # Build the request payload
-    payload: dict[str, Any] = {"description": description}
     if context:
-        # The external service uses user_instruction for additional context
+        description = f"{description}\n\nAdditional context from user:\n{context}"
-        payload["user_instruction"] = context
+
+    payload: dict[str, Any] = {"description": description}
     if library_agents:
         payload["library_agents"] = library_agents
 
@@ -213,24 +212,45 @@ async def decompose_goal_external(
 async def generate_agent_external(
     instructions: dict[str, Any],
     library_agents: list[dict[str, Any]] | None = None,
+    operation_id: str | None = None,
+    task_id: str | None = None,
 ) -> dict[str, Any] | None:
     """Call the external service to generate an agent from instructions.
 
     Args:
         instructions: Structured instructions from decompose_goal
         library_agents: User's library agents available for sub-agent composition
+        operation_id: Operation ID for async processing (enables Redis Streams callback)
+        task_id: Task ID for async processing (enables Redis Streams callback)
 
     Returns:
-        Agent JSON dict on success, or error dict {"type": "error", ...} on error
+        Agent JSON dict, {"status": "accepted"} for async, or error dict {"type": "error", ...} on error
     """
     client = _get_client()
 
+    # Build request payload
     payload: dict[str, Any] = {"instructions": instructions}
     if library_agents:
         payload["library_agents"] = library_agents
+    if operation_id and task_id:
+        payload["operation_id"] = operation_id
+        payload["task_id"] = task_id
 
     try:
         response = await client.post("/api/generate-agent", json=payload)
+
+        # Handle 202 Accepted for async processing
+        if response.status_code == 202:
+            logger.info(
+                f"Agent Generator accepted async request "
+                f"(operation_id={operation_id}, task_id={task_id})"
+            )
+            return {
+                "status": "accepted",
+                "operation_id": operation_id,
+                "task_id": task_id,
+            }
+
         response.raise_for_status()
         data = response.json()
 
@@ -262,6 +282,8 @@ async def generate_agent_patch_external(
     update_request: str,
     current_agent: dict[str, Any],
     library_agents: list[dict[str, Any]] | None = None,
+    operation_id: str | None = None,
+    task_id: str | None = None,
 ) -> dict[str, Any] | None:
     """Call the external service to generate a patch for an existing agent.
 
@@ -269,21 +291,40 @@ async def generate_agent_patch_external(
         update_request: Natural language description of changes
         current_agent: Current agent JSON
         library_agents: User's library agents available for sub-agent composition
+        operation_id: Operation ID for async processing (enables Redis Streams callback)
+        task_id: Task ID for async processing (enables Redis Streams callback)
 
     Returns:
-        Updated agent JSON, clarifying questions dict, or error dict on error
+        Updated agent JSON, clarifying questions dict, {"status": "accepted"} for async, or error dict on error
     """
     client = _get_client()
 
+    # Build request payload
     payload: dict[str, Any] = {
         "update_request": update_request,
         "current_agent_json": current_agent,
     }
     if library_agents:
         payload["library_agents"] = library_agents
+    if operation_id and task_id:
+        payload["operation_id"] = operation_id
+        payload["task_id"] = task_id
 
     try:
         response = await client.post("/api/update-agent", json=payload)
+
+        # Handle 202 Accepted for async processing
+        if response.status_code == 202:
+            logger.info(
+                f"Agent Generator accepted async update request "
+                f"(operation_id={operation_id}, task_id={task_id})"
+            )
+            return {
+                "status": "accepted",
+                "operation_id": operation_id,
+                "task_id": task_id,
+            }
+
         response.raise_for_status()
         data = response.json()
 
@@ -327,6 +368,77 @@ async def generate_agent_patch_external(
         return _create_error_response(error_msg, "unexpected_error")
 
 
+async def customize_template_external(
+    template_agent: dict[str, Any],
+    modification_request: str,
+    context: str = "",
+) -> dict[str, Any] | None:
+    """Call the external service to customize a template/marketplace agent.
+
+    Args:
+        template_agent: The template agent JSON to customize
+        modification_request: Natural language description of customizations
+        context: Additional context (e.g., answers to previous questions)
+
+    Returns:
+        Customized agent JSON, clarifying questions dict, or error dict on error
+    """
+    client = _get_client()
+
+    request = modification_request
+    if context:
+        request = f"{modification_request}\n\nAdditional context from user:\n{context}"
+
+    payload: dict[str, Any] = {
+        "template_agent_json": template_agent,
+        "modification_request": request,
+    }
+
+    try:
+        response = await client.post("/api/template-modification", json=payload)
+        response.raise_for_status()
+        data = response.json()
+
+        if not data.get("success"):
+            error_msg = data.get("error", "Unknown error from Agent Generator")
+            error_type = data.get("error_type", "unknown")
+            logger.error(
+                f"Agent Generator template customization failed: {error_msg} "
+                f"(type: {error_type})"
+            )
+            return _create_error_response(error_msg, error_type)
+
+        # Check if it's clarifying questions
+        if data.get("type") == "clarifying_questions":
+            return {
+                "type": "clarifying_questions",
+                "questions": data.get("questions", []),
+            }
+
+        # Check if it's an error passed through
+        if data.get("type") == "error":
+            return _create_error_response(
+                data.get("error", "Unknown error"),
+                data.get("error_type", "unknown"),
+            )
+
+        # Otherwise return the customized agent JSON
+        return data.get("agent_json")
+
+    except httpx.HTTPStatusError as e:
+        error_type, error_msg = _classify_http_error(e)
+        logger.error(error_msg)
+        return _create_error_response(error_msg, error_type)
+    except httpx.RequestError as e:
+        error_type, error_msg = _classify_request_error(e)
+        logger.error(error_msg)
+        return _create_error_response(error_msg, error_type)
+    except Exception as e:
+        error_msg = f"Unexpected error calling Agent Generator: {e}"
+        logger.error(error_msg)
+        return _create_error_response(error_msg, "unexpected_error")
+
+
 async def get_blocks_external() -> list[dict[str, Any]] | None:
     """Get available blocks from the external service.
 

autogpt_platform/backend/backend/api/features/chat/tools/agent_search.py

@@ -206,9 +206,9 @@ async def search_agents(
             ]
         )
         no_results_msg = (
-            f"No agents found matching '{query}'. Try different keywords or browse the marketplace."
+            f"No agents found matching '{query}'. Let the user know they can try different keywords or browse the marketplace. Also let them know you can create a custom agent for them based on their needs."
             if source == "marketplace"
-            else f"No agents matching '{query}' found in your library."
+            else f"No agents matching '{query}' found in your library. Let the user know you can create a custom agent for them based on their needs."
         )
         return NoResultsResponse(
             message=no_results_msg, session_id=session_id, suggestions=suggestions
@@ -224,10 +224,10 @@ async def search_agents(
     message = (
         "Now you have found some options for the user to choose from. "
         "You can add a link to a recommended agent at: /marketplace/agent/agent_id "
-        "Please ask the user if they would like to use any of these agents."
+        "Please ask the user if they would like to use any of these agents. Let the user know we can create a custom agent for them based on their needs."
         if source == "marketplace"
         else "Found agents in the user's library. You can provide a link to view an agent at: "
-        "/library/agents/{agent_id}. Use agent_output to get execution results, or run_agent to execute."
+        "/library/agents/{agent_id}. Use agent_output to get execution results, or run_agent to execute. Let the user know we can create a custom agent for them based on their needs."
     )
 
     return AgentsFoundResponse(

autogpt_platform/backend/backend/api/features/chat/tools/create_agent.py

@@ -18,6 +18,7 @@ from .base import BaseTool
 from .models import (
     AgentPreviewResponse,
     AgentSavedResponse,
+    AsyncProcessingResponse,
     ClarificationNeededResponse,
     ClarifyingQuestion,
     ErrorResponse,
@@ -98,6 +99,10 @@ class CreateAgentTool(BaseTool):
         save = kwargs.get("save", True)
         session_id = session.session_id if session else None
 
+        # Extract async processing params (passed by long-running tool handler)
+        operation_id = kwargs.get("_operation_id")
+        task_id = kwargs.get("_task_id")
+
         if not description:
             return ErrorResponse(
                 message="Please provide a description of what the agent should do.",
@@ -219,7 +224,12 @@ class CreateAgentTool(BaseTool):
                 logger.warning(f"Failed to enrich library agents from steps: {e}")
 
         try:
-            agent_json = await generate_agent(decomposition_result, library_agents)
+            agent_json = await generate_agent(
+                decomposition_result,
+                library_agents,
+                operation_id=operation_id,
+                task_id=task_id,
+            )
         except AgentGeneratorNotConfiguredError:
             return ErrorResponse(
                 message=(
@@ -263,6 +273,19 @@ class CreateAgentTool(BaseTool):
                 session_id=session_id,
             )
 
+        # Check if Agent Generator accepted for async processing
+        if agent_json.get("status") == "accepted":
+            logger.info(
+                f"Agent generation delegated to async processing "
+                f"(operation_id={operation_id}, task_id={task_id})"
+            )
+            return AsyncProcessingResponse(
+                message="Agent generation started. You'll be notified when it's complete.",
+                operation_id=operation_id,
+                task_id=task_id,
+                session_id=session_id,
+            )
+
         agent_name = agent_json.get("name", "Generated Agent")
         agent_description = agent_json.get("description", "")
         node_count = len(agent_json.get("nodes", []))

autogpt_platform/backend/backend/api/features/chat/tools/customize_agent.py

@@ -0,0 +1,337 @@
+"""CustomizeAgentTool - Customizes marketplace/template agents using natural language."""
+
+import logging
+from typing import Any
+
+from backend.api.features.chat.model import ChatSession
+from backend.api.features.store import db as store_db
+from backend.api.features.store.exceptions import AgentNotFoundError
+
+from .agent_generator import (
+    AgentGeneratorNotConfiguredError,
+    customize_template,
+    get_user_message_for_error,
+    graph_to_json,
+    save_agent_to_library,
+)
+from .base import BaseTool
+from .models import (
+    AgentPreviewResponse,
+    AgentSavedResponse,
+    ClarificationNeededResponse,
+    ClarifyingQuestion,
+    ErrorResponse,
+    ToolResponseBase,
+)
+
+logger = logging.getLogger(__name__)
+
+
+class CustomizeAgentTool(BaseTool):
+    """Tool for customizing marketplace/template agents using natural language."""
+
+    @property
+    def name(self) -> str:
+        return "customize_agent"
+
+    @property
+    def description(self) -> str:
+        return (
+            "Customize a marketplace or template agent using natural language. "
+            "Takes an existing agent from the marketplace and modifies it based on "
+            "the user's requirements before adding to their library."
+        )
+
+    @property
+    def requires_auth(self) -> bool:
+        return True
+
+    @property
+    def is_long_running(self) -> bool:
+        return True
+
+    @property
+    def parameters(self) -> dict[str, Any]:
+        return {
+            "type": "object",
+            "properties": {
+                "agent_id": {
+                    "type": "string",
+                    "description": (
+                        "The marketplace agent ID in format 'creator/slug' "
+                        "(e.g., 'autogpt/newsletter-writer'). "
+                        "Get this from find_agent results."
+                    ),
+                },
+                "modifications": {
+                    "type": "string",
+                    "description": (
+                        "Natural language description of how to customize the agent. "
+                        "Be specific about what changes you want to make."
+                    ),
+                },
+                "context": {
+                    "type": "string",
+                    "description": (
+                        "Additional context or answers to previous clarifying questions."
+                    ),
+                },
+                "save": {
+                    "type": "boolean",
+                    "description": (
+                        "Whether to save the customized agent to the user's library. "
+                        "Default is true. Set to false for preview only."
+                    ),
+                    "default": True,
+                },
+            },
+            "required": ["agent_id", "modifications"],
+        }
+
+    async def _execute(
+        self,
+        user_id: str | None,
+        session: ChatSession,
+        **kwargs,
+    ) -> ToolResponseBase:
+        """Execute the customize_agent tool.
+
+        Flow:
+        1. Parse the agent ID to get creator/slug
+        2. Fetch the template agent from the marketplace
+        3. Call customize_template with the modification request
+        4. Preview or save based on the save parameter
+        """
+        agent_id = kwargs.get("agent_id", "").strip()
+        modifications = kwargs.get("modifications", "").strip()
+        context = kwargs.get("context", "")
+        save = kwargs.get("save", True)
+        session_id = session.session_id if session else None
+
+        if not agent_id:
+            return ErrorResponse(
+                message="Please provide the marketplace agent ID (e.g., 'creator/agent-name').",
+                error="missing_agent_id",
+                session_id=session_id,
+            )
+
+        if not modifications:
+            return ErrorResponse(
+                message="Please describe how you want to customize this agent.",
+                error="missing_modifications",
+                session_id=session_id,
+            )
+
+        # Parse agent_id in format "creator/slug"
+        parts = [p.strip() for p in agent_id.split("/")]
+        if len(parts) != 2 or not parts[0] or not parts[1]:
+            return ErrorResponse(
+                message=(
+                    f"Invalid agent ID format: '{agent_id}'. "
+                    "Expected format is 'creator/agent-name' "
+                    "(e.g., 'autogpt/newsletter-writer')."
+                ),
+                error="invalid_agent_id_format",
+                session_id=session_id,
+            )
+
+        creator_username, agent_slug = parts
+
+        # Fetch the marketplace agent details
+        try:
+            agent_details = await store_db.get_store_agent_details(
+                username=creator_username, agent_name=agent_slug
+            )
+        except AgentNotFoundError:
+            return ErrorResponse(
+                message=(
+                    f"Could not find marketplace agent '{agent_id}'. "
+                    "Please check the agent ID and try again."
+                ),
+                error="agent_not_found",
+                session_id=session_id,
+            )
+        except Exception as e:
+            logger.error(f"Error fetching marketplace agent {agent_id}: {e}")
+            return ErrorResponse(
+                message="Failed to fetch the marketplace agent. Please try again.",
+                error="fetch_error",
+                session_id=session_id,
+            )
+
+        if not agent_details.store_listing_version_id:
+            return ErrorResponse(
+                message=(
+                    f"The agent '{agent_id}' does not have an available version. "
+                    "Please try a different agent."
+                ),
+                error="no_version_available",
+                session_id=session_id,
+            )
+
+        # Get the full agent graph
+        try:
+            graph = await store_db.get_agent(agent_details.store_listing_version_id)
+            template_agent = graph_to_json(graph)
+        except Exception as e:
+            logger.error(f"Error fetching agent graph for {agent_id}: {e}")
+            return ErrorResponse(
+                message="Failed to fetch the agent configuration. Please try again.",
+                error="graph_fetch_error",
+                session_id=session_id,
+            )
+
+        # Call customize_template
+        try:
+            result = await customize_template(
+                template_agent=template_agent,
+                modification_request=modifications,
+                context=context,
+            )
+        except AgentGeneratorNotConfiguredError:
+            return ErrorResponse(
+                message=(
+                    "Agent customization is not available. "
+                    "The Agent Generator service is not configured."
+                ),
+                error="service_not_configured",
+                session_id=session_id,
+            )
+        except Exception as e:
+            logger.error(f"Error calling customize_template for {agent_id}: {e}")
+            return ErrorResponse(
+                message=(
+                    "Failed to customize the agent due to a service error. "
+                    "Please try again."
+                ),
+                error="customization_service_error",
+                session_id=session_id,
+            )
+
+        if result is None:
+            return ErrorResponse(
+                message=(
+                    "Failed to customize the agent. "
+                    "The agent generation service may be unavailable or timed out. "
+                    "Please try again."
+                ),
+                error="customization_failed",
+                session_id=session_id,
+            )
+
+        # Handle error response
+        if isinstance(result, dict) and result.get("type") == "error":
+            error_msg = result.get("error", "Unknown error")
+            error_type = result.get("error_type", "unknown")
+            user_message = get_user_message_for_error(
+                error_type,
+                operation="customize the agent",
+                llm_parse_message=(
+                    "The AI had trouble customizing the agent. "
+                    "Please try again or simplify your request."
+                ),
+                validation_message=(
+                    "The customized agent failed validation. "
+                    "Please try rephrasing your request."
+                ),
+                error_details=error_msg,
+            )
+            return ErrorResponse(
+                message=user_message,
+                error=f"customization_failed:{error_type}",
+                session_id=session_id,
+            )
+
+        # Handle clarifying questions
+        if isinstance(result, dict) and result.get("type") == "clarifying_questions":
+            questions = result.get("questions") or []
+            if not isinstance(questions, list):
+                logger.error(
+                    f"Unexpected clarifying questions format: {type(questions)}"
+                )
+                questions = []
+            return ClarificationNeededResponse(
+                message=(
+                    "I need some more information to customize this agent. "
+                    "Please answer the following questions:"
+                ),
+                questions=[
+                    ClarifyingQuestion(
+                        question=q.get("question", ""),
+                        keyword=q.get("keyword", ""),
+                        example=q.get("example"),
+                    )
+                    for q in questions
+                    if isinstance(q, dict)
+                ],
+                session_id=session_id,
+            )
+
+        # Result should be the customized agent JSON
+        if not isinstance(result, dict):
+            logger.error(f"Unexpected customize_template response type: {type(result)}")
+            return ErrorResponse(
+                message="Failed to customize the agent due to an unexpected response.",
+                error="unexpected_response_type",
+                session_id=session_id,
+            )
+
+        customized_agent = result
+
+        agent_name = customized_agent.get(
+            "name", f"Customized {agent_details.agent_name}"
+        )
+        agent_description = customized_agent.get("description", "")
+        nodes = customized_agent.get("nodes")
+        links = customized_agent.get("links")
+        node_count = len(nodes) if isinstance(nodes, list) else 0
+        link_count = len(links) if isinstance(links, list) else 0
+
+        if not save:
+            return AgentPreviewResponse(
+                message=(
+                    f"I've customized the agent '{agent_details.agent_name}'. "
+                    f"The customized agent has {node_count} blocks. "
+                    f"Review it and call customize_agent with save=true to save it."
+                ),
+                agent_json=customized_agent,
+                agent_name=agent_name,
+                description=agent_description,
+                node_count=node_count,
+                link_count=link_count,
+                session_id=session_id,
+            )
+
+        if not user_id:
+            return ErrorResponse(
+                message="You must be logged in to save agents.",
+                error="auth_required",
+                session_id=session_id,
+            )
+
+        # Save to user's library
+        try:
+            created_graph, library_agent = await save_agent_to_library(
+                customized_agent, user_id, is_update=False
+            )
+
+            return AgentSavedResponse(
+                message=(
+                    f"Customized agent '{created_graph.name}' "
+                    f"(based on '{agent_details.agent_name}') "
+                    f"has been saved to your library!"
+                ),
+                agent_id=created_graph.id,
+                agent_name=created_graph.name,
+                library_agent_id=library_agent.id,
+                library_agent_link=f"/library/agents/{library_agent.id}",
+                agent_page_link=f"/build?flowID={created_graph.id}",
+                session_id=session_id,
+            )
+        except Exception as e:
+            logger.error(f"Error saving customized agent: {e}")
+            return ErrorResponse(
+                message="Failed to save the customized agent. Please try again.",
+                error="save_failed",
+                session_id=session_id,
+            )

autogpt_platform/backend/backend/api/features/chat/tools/edit_agent.py

@@ -17,6 +17,7 @@ from .base import BaseTool
 from .models import (
     AgentPreviewResponse,
     AgentSavedResponse,
+    AsyncProcessingResponse,
     ClarificationNeededResponse,
     ClarifyingQuestion,
     ErrorResponse,
@@ -104,6 +105,10 @@ class EditAgentTool(BaseTool):
         save = kwargs.get("save", True)
         session_id = session.session_id if session else None
 
+        # Extract async processing params (passed by long-running tool handler)
+        operation_id = kwargs.get("_operation_id")
+        task_id = kwargs.get("_task_id")
+
         if not agent_id:
             return ErrorResponse(
                 message="Please provide the agent ID to edit.",
@@ -149,7 +154,11 @@ class EditAgentTool(BaseTool):
 
         try:
             result = await generate_agent_patch(
-                update_request, current_agent, library_agents
+                update_request,
+                current_agent,
+                library_agents,
+                operation_id=operation_id,
+                task_id=task_id,
             )
         except AgentGeneratorNotConfiguredError:
             return ErrorResponse(
@@ -169,6 +178,20 @@ class EditAgentTool(BaseTool):
                 session_id=session_id,
             )
 
+        # Check if Agent Generator accepted for async processing
+        if result.get("status") == "accepted":
+            logger.info(
+                f"Agent edit delegated to async processing "
+                f"(operation_id={operation_id}, task_id={task_id})"
+            )
+            return AsyncProcessingResponse(
+                message="Agent edit started. You'll be notified when it's complete.",
+                operation_id=operation_id,
+                task_id=task_id,
+                session_id=session_id,
+            )
+
+        # Check if the result is an error from the external service
         if isinstance(result, dict) and result.get("type") == "error":
             error_msg = result.get("error", "Unknown error")
             error_type = result.get("error_type", "unknown")

autogpt_platform/backend/backend/api/features/chat/tools/models.py

@@ -38,6 +38,8 @@ class ResponseType(str, Enum):
     OPERATION_STARTED = "operation_started"
     OPERATION_PENDING = "operation_pending"
     OPERATION_IN_PROGRESS = "operation_in_progress"
+    # Input validation
+    INPUT_VALIDATION_ERROR = "input_validation_error"
 
 
 # Base response model
@@ -68,6 +70,10 @@ class AgentInfo(BaseModel):
     has_external_trigger: bool | None = None
     new_output: bool | None = None
     graph_id: str | None = None
+    inputs: dict[str, Any] | None = Field(
+        default=None,
+        description="Input schema for the agent, including field names, types, and defaults",
+    )
 
 
 class AgentsFoundResponse(ToolResponseBase):
@@ -194,6 +200,20 @@ class ErrorResponse(ToolResponseBase):
     details: dict[str, Any] | None = None
 
 
+class InputValidationErrorResponse(ToolResponseBase):
+    """Response when run_agent receives unknown input fields."""
+
+    type: ResponseType = ResponseType.INPUT_VALIDATION_ERROR
+    unrecognized_fields: list[str] = Field(
+        description="List of input field names that were not recognized"
+    )
+    inputs: dict[str, Any] = Field(
+        description="The agent's valid input schema for reference"
+    )
+    graph_id: str | None = None
+    graph_version: int | None = None
+
+
 # Agent output models
 class ExecutionOutputInfo(BaseModel):
     """Summary of a single execution's outputs."""
@@ -352,11 +372,15 @@ class OperationStartedResponse(ToolResponseBase):
 
     This is returned immediately to the client while the operation continues
     to execute. The user can close the tab and check back later.
+
+    The task_id can be used to reconnect to the SSE stream via
+    GET /chat/tasks/{task_id}/stream?last_idx=0
     """
 
     type: ResponseType = ResponseType.OPERATION_STARTED
     operation_id: str
     tool_name: str
+    task_id: str | None = None  # For SSE reconnection
 
 
 class OperationPendingResponse(ToolResponseBase):
@@ -380,3 +404,20 @@ class OperationInProgressResponse(ToolResponseBase):
 
     type: ResponseType = ResponseType.OPERATION_IN_PROGRESS
     tool_call_id: str
+
+
+class AsyncProcessingResponse(ToolResponseBase):
+    """Response when an operation has been delegated to async processing.
+
+    This is returned by tools when the external service accepts the request
+    for async processing (HTTP 202 Accepted). The Redis Streams completion
+    consumer will handle the result when the external service completes.
+
+    The status field is specifically "accepted" to allow the long-running tool
+    handler to detect this response and skip LLM continuation.
+    """
+
+    type: ResponseType = ResponseType.OPERATION_STARTED
+    status: str = "accepted"  # Must be "accepted" for detection
+    operation_id: str | None = None
+    task_id: str | None = None

autogpt_platform/backend/backend/api/features/chat/tools/run_agent.py

@@ -30,6 +30,7 @@ from .models import (
     ErrorResponse,
     ExecutionOptions,
     ExecutionStartedResponse,
+    InputValidationErrorResponse,
     SetupInfo,
     SetupRequirementsResponse,
     ToolResponseBase,
@@ -273,6 +274,22 @@ class RunAgentTool(BaseTool):
             input_properties = graph.input_schema.get("properties", {})
             required_fields = set(graph.input_schema.get("required", []))
             provided_inputs = set(params.inputs.keys())
+            valid_fields = set(input_properties.keys())
+
+            # Check for unknown input fields
+            unrecognized_fields = provided_inputs - valid_fields
+            if unrecognized_fields:
+                return InputValidationErrorResponse(
+                    message=(
+                        f"Unknown input field(s) provided: {', '.join(sorted(unrecognized_fields))}. "
+                        f"Agent was not executed. Please use the correct field names from the schema."
+                    ),
+                    session_id=session_id,
+                    unrecognized_fields=sorted(unrecognized_fields),
+                    inputs=graph.input_schema,
+                    graph_id=graph.id,
+                    graph_version=graph.version,
+                )
 
             # If agent has inputs but none were provided AND use_defaults is not set,
             # always show what's available first so user can decide

autogpt_platform/backend/backend/api/features/chat/tools/run_agent_test.py

@@ -402,3 +402,42 @@ async def test_run_agent_schedule_without_name(setup_test_data):
     # Should return error about missing schedule_name
     assert result_data.get("type") == "error"
     assert "schedule_name" in result_data["message"].lower()
+
+
+@pytest.mark.asyncio(loop_scope="session")
+async def test_run_agent_rejects_unknown_input_fields(setup_test_data):
+    """Test that run_agent returns input_validation_error for unknown input fields."""
+    user = setup_test_data["user"]
+    store_submission = setup_test_data["store_submission"]
+
+    tool = RunAgentTool()
+    agent_marketplace_id = f"{user.email.split('@')[0]}/{store_submission.slug}"
+    session = make_session(user_id=user.id)
+
+    # Execute with unknown input field names
+    response = await tool.execute(
+        user_id=user.id,
+        session_id=str(uuid.uuid4()),
+        tool_call_id=str(uuid.uuid4()),
+        username_agent_slug=agent_marketplace_id,
+        inputs={
+            "unknown_field": "some value",
+            "another_unknown": "another value",
+        },
+        session=session,
+    )
+
+    assert response is not None
+    assert hasattr(response, "output")
+    assert isinstance(response.output, str)
+    result_data = orjson.loads(response.output)
+
+    # Should return input_validation_error type with unrecognized fields
+    assert result_data.get("type") == "input_validation_error"
+    assert "unrecognized_fields" in result_data
+    assert set(result_data["unrecognized_fields"]) == {
+        "another_unknown",
+        "unknown_field",
+    }
+    assert "inputs" in result_data  # Contains the valid schema
+    assert "Agent was not executed" in result_data["message"]

autogpt_platform/backend/backend/api/features/chat/tools/run_block.py

@@ -5,6 +5,8 @@ import uuid
 from collections import defaultdict
 from typing import Any
 
+from pydantic_core import PydanticUndefined
+
 from backend.api.features.chat.model import ChatSession
 from backend.data.block import get_block
 from backend.data.execution import ExecutionContext
@@ -75,15 +77,22 @@ class RunBlockTool(BaseTool):
         self,
         user_id: str,
         block: Any,
+        input_data: dict[str, Any] | None = None,
     ) -> tuple[dict[str, CredentialsMetaInput], list[CredentialsMetaInput]]:
         """
         Check if user has required credentials for a block.
 
+        Args:
+            user_id: User ID
+            block: Block to check credentials for
+            input_data: Input data for the block (used to determine provider via discriminator)
+
         Returns:
             tuple[matched_credentials, missing_credentials]
         """
         matched_credentials: dict[str, CredentialsMetaInput] = {}
         missing_credentials: list[CredentialsMetaInput] = []
+        input_data = input_data or {}
 
         # Get credential field info from block's input schema
         credentials_fields_info = block.input_schema.get_credentials_fields_info()
@@ -96,14 +105,33 @@ class RunBlockTool(BaseTool):
         available_creds = await creds_manager.store.get_all_creds(user_id)
 
         for field_name, field_info in credentials_fields_info.items():
-            # field_info.provider is a frozenset of acceptable providers
+            effective_field_info = field_info
-            # field_info.supported_types is a frozenset of acceptable types
+            if field_info.discriminator and field_info.discriminator_mapping:
+                # Get discriminator from input, falling back to schema default
+                discriminator_value = input_data.get(field_info.discriminator)
+                if discriminator_value is None:
+                    field = block.input_schema.model_fields.get(
+                        field_info.discriminator
+                    )
+                    if field and field.default is not PydanticUndefined:
+                        discriminator_value = field.default
+
+                if (
+                    discriminator_value
+                    and discriminator_value in field_info.discriminator_mapping
+                ):
+                    effective_field_info = field_info.discriminate(discriminator_value)
+                    logger.debug(
+                        f"Discriminated provider for {field_name}: "
+                        f"{discriminator_value} -> {effective_field_info.provider}"
+                    )
+
             matching_cred = next(
                 (
                     cred
                     for cred in available_creds
-                    if cred.provider in field_info.provider
+                    if cred.provider in effective_field_info.provider
-                    and cred.type in field_info.supported_types
+                    and cred.type in effective_field_info.supported_types
                 ),
                 None,
             )
@@ -117,8 +145,8 @@ class RunBlockTool(BaseTool):
                 )
             else:
                 # Create a placeholder for the missing credential
-                provider = next(iter(field_info.provider), "unknown")
+                provider = next(iter(effective_field_info.provider), "unknown")
-                cred_type = next(iter(field_info.supported_types), "api_key")
+                cred_type = next(iter(effective_field_info.supported_types), "api_key")
                 missing_credentials.append(
                     CredentialsMetaInput(
                         id=field_name,
@@ -186,10 +214,9 @@ class RunBlockTool(BaseTool):
 
         logger.info(f"Executing block {block.name} ({block_id}) for user {user_id}")
 
-        # Check credentials
         creds_manager = IntegrationCredentialsManager()
         matched_credentials, missing_credentials = await self._check_block_credentials(
-            user_id, block
+            user_id, block, input_data
         )
 
         if missing_credentials:

autogpt_platform/backend/backend/api/features/chat/tools/utils.py

@@ -8,7 +8,12 @@ from backend.api.features.library import model as library_model
 from backend.api.features.store import db as store_db
 from backend.data import graph as graph_db
 from backend.data.graph import GraphModel
-from backend.data.model import Credentials, CredentialsFieldInfo, CredentialsMetaInput
+from backend.data.model import (
+    CredentialsFieldInfo,
+    CredentialsMetaInput,
+    HostScopedCredentials,
+    OAuth2Credentials,
+)
 from backend.integrations.creds_manager import IntegrationCredentialsManager
 from backend.util.exceptions import NotFoundError
 
@@ -273,7 +278,14 @@ async def match_user_credentials_to_graph(
                 for cred in available_creds
                 if cred.provider in credential_requirements.provider
                 and cred.type in credential_requirements.supported_types
-                and _credential_has_required_scopes(cred, credential_requirements)
+                and (
+                    cred.type != "oauth2"
+                    or _credential_has_required_scopes(cred, credential_requirements)
+                )
+                and (
+                    cred.type != "host_scoped"
+                    or _credential_is_for_host(cred, credential_requirements)
+                )
             ),
             None,
         )
@@ -318,19 +330,10 @@ async def match_user_credentials_to_graph(
 
 
 def _credential_has_required_scopes(
-    credential: Credentials,
+    credential: OAuth2Credentials,
     requirements: CredentialsFieldInfo,
 ) -> bool:
-    """
+    """Check if an OAuth2 credential has all the scopes required by the input."""
-    Check if a credential has all the scopes required by the block.
-
-    For OAuth2 credentials, verifies that the credential's scopes are a superset
-    of the required scopes. For other credential types, returns True (no scope check).
-    """
-    # Only OAuth2 credentials have scopes to check
-    if credential.type != "oauth2":
-        return True
-
     # If no scopes are required, any credential matches
     if not requirements.required_scopes:
         return True
@@ -339,6 +342,22 @@ def _credential_has_required_scopes(
     return set(credential.scopes).issuperset(requirements.required_scopes)
 
 
+def _credential_is_for_host(
+    credential: HostScopedCredentials,
+    requirements: CredentialsFieldInfo,
+) -> bool:
+    """Check if a host-scoped credential matches the host required by the input."""
+    # We need to know the host to match host-scoped credentials to.
+    # Graph.aggregate_credentials_inputs() adds the node's set URL value (if any)
+    # to discriminator_values. No discriminator_values -> no host to match against.
+    if not requirements.discriminator_values:
+        return True
+
+    # Check that credential host matches required host.
+    # Host-scoped credential inputs are grouped by host, so any item from the set works.
+    return credential.matches_url(list(requirements.discriminator_values)[0])
+
+
 async def check_user_has_required_credentials(
     user_id: str,
     required_credentials: list[CredentialsMetaInput],

autogpt_platform/backend/backend/api/features/library/db.py

@@ -19,7 +19,10 @@ from backend.data.graph import GraphSettings
 from backend.data.includes import AGENT_PRESET_INCLUDE, library_agent_include
 from backend.data.model import CredentialsMetaInput
 from backend.integrations.creds_manager import IntegrationCredentialsManager
-from backend.integrations.webhooks.graph_lifecycle_hooks import on_graph_activate
+from backend.integrations.webhooks.graph_lifecycle_hooks import (
+    on_graph_activate,
+    on_graph_deactivate,
+)
 from backend.util.clients import get_scheduler_client
 from backend.util.exceptions import DatabaseError, InvalidInputError, NotFoundError
 from backend.util.json import SafeJson
@@ -537,6 +540,92 @@ async def update_agent_version_in_library(
     return library_model.LibraryAgent.from_db(lib)
 
 
+async def create_graph_in_library(
+    graph: graph_db.Graph,
+    user_id: str,
+) -> tuple[graph_db.GraphModel, library_model.LibraryAgent]:
+    """Create a new graph and add it to the user's library."""
+    graph.version = 1
+    graph_model = graph_db.make_graph_model(graph, user_id)
+    graph_model.reassign_ids(user_id=user_id, reassign_graph_id=True)
+
+    created_graph = await graph_db.create_graph(graph_model, user_id)
+
+    library_agents = await create_library_agent(
+        graph=created_graph,
+        user_id=user_id,
+        sensitive_action_safe_mode=True,
+        create_library_agents_for_sub_graphs=False,
+    )
+
+    if created_graph.is_active:
+        created_graph = await on_graph_activate(created_graph, user_id=user_id)
+
+    return created_graph, library_agents[0]
+
+
+async def update_graph_in_library(
+    graph: graph_db.Graph,
+    user_id: str,
+) -> tuple[graph_db.GraphModel, library_model.LibraryAgent]:
+    """Create a new version of an existing graph and update the library entry."""
+    existing_versions = await graph_db.get_graph_all_versions(graph.id, user_id)
+    current_active_version = (
+        next((v for v in existing_versions if v.is_active), None)
+        if existing_versions
+        else None
+    )
+    graph.version = (
+        max(v.version for v in existing_versions) + 1 if existing_versions else 1
+    )
+
+    graph_model = graph_db.make_graph_model(graph, user_id)
+    graph_model.reassign_ids(user_id=user_id, reassign_graph_id=False)
+
+    created_graph = await graph_db.create_graph(graph_model, user_id)
+
+    library_agent = await get_library_agent_by_graph_id(user_id, created_graph.id)
+    if not library_agent:
+        raise NotFoundError(f"Library agent not found for graph {created_graph.id}")
+
+    library_agent = await update_library_agent_version_and_settings(
+        user_id, created_graph
+    )
+
+    if created_graph.is_active:
+        created_graph = await on_graph_activate(created_graph, user_id=user_id)
+        await graph_db.set_graph_active_version(
+            graph_id=created_graph.id,
+            version=created_graph.version,
+            user_id=user_id,
+        )
+        if current_active_version:
+            await on_graph_deactivate(current_active_version, user_id=user_id)
+
+    return created_graph, library_agent
+
+
+async def update_library_agent_version_and_settings(
+    user_id: str, agent_graph: graph_db.GraphModel
+) -> library_model.LibraryAgent:
+    """Update library agent to point to new graph version and sync settings."""
+    library = await update_agent_version_in_library(
+        user_id, agent_graph.id, agent_graph.version
+    )
+    updated_settings = GraphSettings.from_graph(
+        graph=agent_graph,
+        hitl_safe_mode=library.settings.human_in_the_loop_safe_mode,
+        sensitive_action_safe_mode=library.settings.sensitive_action_safe_mode,
+    )
+    if updated_settings != library.settings:
+        library = await update_library_agent(
+            library_agent_id=library.id,
+            user_id=user_id,
+            settings=updated_settings,
+        )
+    return library
+
+
 async def update_library_agent(
     library_agent_id: str,
     user_id: str,

autogpt_platform/backend/backend/api/features/store/db.py

@@ -112,6 +112,7 @@ async def get_store_agents(
                             description=agent["description"],
                             runs=agent["runs"],
                             rating=agent["rating"],
+                            agent_graph_id=agent.get("agentGraphId", ""),
                         )
                         store_agents.append(store_agent)
                     except Exception as e:
@@ -170,6 +171,7 @@ async def get_store_agents(
                         description=agent.description,
                         runs=agent.runs,
                         rating=agent.rating,
+                        agent_graph_id=agent.agentGraphId,
                     )
                     # Add to the list only if creation was successful
                     store_agents.append(store_agent)

autogpt_platform/backend/backend/api/features/store/embeddings_e2e_test.py

@@ -454,6 +454,9 @@ async def test_unified_hybrid_search_pagination(
     cleanup_embeddings: list,
 ):
     """Test unified search pagination works correctly."""
+    # Use a unique search term to avoid matching other test data
+    unique_term = f"xyzpagtest{uuid.uuid4().hex[:8]}"
+
     # Create multiple items
     content_ids = []
     for i in range(5):
@@ -465,14 +468,14 @@ async def test_unified_hybrid_search_pagination(
             content_type=ContentType.BLOCK,
             content_id=content_id,
             embedding=mock_embedding,
-            searchable_text=f"pagination test item number {i}",
+            searchable_text=f"{unique_term} item number {i}",
             metadata={"index": i},
             user_id=None,
         )
 
     # Get first page
     page1_results, total1 = await unified_hybrid_search(
-        query="pagination test",
+        query=unique_term,
         content_types=[ContentType.BLOCK],
         page=1,
         page_size=2,
@@ -480,7 +483,7 @@ async def test_unified_hybrid_search_pagination(
 
     # Get second page
     page2_results, total2 = await unified_hybrid_search(
-        query="pagination test",
+        query=unique_term,
         content_types=[ContentType.BLOCK],
         page=2,
         page_size=2,

autogpt_platform/backend/backend/api/features/store/hybrid_search.py

@@ -600,6 +600,7 @@ async def hybrid_search(
                 sa.featured,
                 sa.is_available,
                 sa.updated_at,
+                sa."agentGraphId",
                 -- Searchable text for BM25 reranking
                 COALESCE(sa.agent_name, '') || ' ' || COALESCE(sa.sub_heading, '') || ' ' || COALESCE(sa.description, '') as searchable_text,
                 -- Semantic score
@@ -659,6 +660,7 @@ async def hybrid_search(
                 featured,
                 is_available,
                 updated_at,
+                "agentGraphId",
                 searchable_text,
                 semantic_score,
                 lexical_score,

autogpt_platform/backend/backend/api/features/store/model.py

@@ -38,6 +38,7 @@ class StoreAgent(pydantic.BaseModel):
     description: str
     runs: int
     rating: float
+    agent_graph_id: str
 
 
 class StoreAgentsResponse(pydantic.BaseModel):

autogpt_platform/backend/backend/api/features/store/model_test.py

@@ -26,11 +26,13 @@ def test_store_agent():
         description="Test description",
         runs=50,
         rating=4.5,
+        agent_graph_id="test-graph-id",
     )
     assert agent.slug == "test-agent"
     assert agent.agent_name == "Test Agent"
     assert agent.runs == 50
     assert agent.rating == 4.5
+    assert agent.agent_graph_id == "test-graph-id"
 
 
 def test_store_agents_response():
@@ -46,6 +48,7 @@ def test_store_agents_response():
                 description="Test description",
                 runs=50,
                 rating=4.5,
+                agent_graph_id="test-graph-id",
             )
         ],
         pagination=store_model.Pagination(

autogpt_platform/backend/backend/api/features/store/routes_test.py

@@ -82,6 +82,7 @@ def test_get_agents_featured(
                 description="Featured agent description",
                 runs=100,
                 rating=4.5,
+                agent_graph_id="test-graph-1",
             )
         ],
         pagination=store_model.Pagination(
@@ -127,6 +128,7 @@ def test_get_agents_by_creator(
                 description="Creator agent description",
                 runs=50,
                 rating=4.0,
+                agent_graph_id="test-graph-2",
             )
         ],
         pagination=store_model.Pagination(
@@ -172,6 +174,7 @@ def test_get_agents_sorted(
                 description="Top agent description",
                 runs=1000,
                 rating=5.0,
+                agent_graph_id="test-graph-3",
             )
         ],
         pagination=store_model.Pagination(
@@ -217,6 +220,7 @@ def test_get_agents_search(
                 description="Specific search term description",
                 runs=75,
                 rating=4.2,
+                agent_graph_id="test-graph-search",
             )
         ],
         pagination=store_model.Pagination(
@@ -262,6 +266,7 @@ def test_get_agents_category(
                 description="Category agent description",
                 runs=60,
                 rating=4.1,
+                agent_graph_id="test-graph-category",
             )
         ],
         pagination=store_model.Pagination(
@@ -306,6 +311,7 @@ def test_get_agents_pagination(
                 description=f"Agent {i} description",
                 runs=i * 10,
                 rating=4.0,
+                agent_graph_id="test-graph-2",
             )
             for i in range(5)
         ],

autogpt_platform/backend/backend/api/features/store/test_cache_delete.py

@@ -33,6 +33,7 @@ class TestCacheDeletion:
                     description="Test description",
                     runs=100,
                     rating=4.5,
+                    agent_graph_id="test-graph-id",
                 )
             ],
             pagination=Pagination(

autogpt_platform/backend/backend/api/features/v1.py

@@ -101,7 +101,6 @@ from backend.util.timezone_utils import (
 from backend.util.virus_scanner import scan_content_safe
 
 from .library import db as library_db
-from .library import model as library_model
 from .store.model import StoreAgentDetails
 
 
@@ -823,18 +822,16 @@ async def update_graph(
     graph: graph_db.Graph,
     user_id: Annotated[str, Security(get_user_id)],
 ) -> graph_db.GraphModel:
-    # Sanity check
     if graph.id and graph.id != graph_id:
         raise HTTPException(400, detail="Graph ID does not match ID in URI")
 
-    # Determine new version
     existing_versions = await graph_db.get_graph_all_versions(graph_id, user_id=user_id)
     if not existing_versions:
         raise HTTPException(404, detail=f"Graph #{graph_id} not found")
-    latest_version_number = max(g.version for g in existing_versions)
-    graph.version = latest_version_number + 1
 
+    graph.version = max(g.version for g in existing_versions) + 1
     current_active_version = next((v for v in existing_versions if v.is_active), None)
+
     graph = graph_db.make_graph_model(graph, user_id)
     graph.reassign_ids(user_id=user_id, reassign_graph_id=False)
     graph.validate_graph(for_run=False)
@@ -842,27 +839,23 @@ async def update_graph(
     new_graph_version = await graph_db.create_graph(graph, user_id=user_id)
 
     if new_graph_version.is_active:
-        # Keep the library agent up to date with the new active version
+        await library_db.update_library_agent_version_and_settings(
-        await _update_library_agent_version_and_settings(user_id, new_graph_version)
+            user_id, new_graph_version
-
+        )
-        # Handle activation of the new graph first to ensure continuity
         new_graph_version = await on_graph_activate(new_graph_version, user_id=user_id)
-        # Ensure new version is the only active version
         await graph_db.set_graph_active_version(
             graph_id=graph_id, version=new_graph_version.version, user_id=user_id
         )
         if current_active_version:
-            # Handle deactivation of the previously active version
             await on_graph_deactivate(current_active_version, user_id=user_id)
 
-    # Fetch new graph version *with sub-graphs* (needed for credentials input schema)
     new_graph_version_with_subgraphs = await graph_db.get_graph(
         graph_id,
         new_graph_version.version,
         user_id=user_id,
         include_subgraphs=True,
     )
-    assert new_graph_version_with_subgraphs  # make type checker happy
+    assert new_graph_version_with_subgraphs
     return new_graph_version_with_subgraphs
 
 
@@ -900,33 +893,15 @@ async def set_graph_active_version(
     )
 
     # Keep the library agent up to date with the new active version
-    await _update_library_agent_version_and_settings(user_id, new_active_graph)
+    await library_db.update_library_agent_version_and_settings(
+        user_id, new_active_graph
+    )
 
     if current_active_graph and current_active_graph.version != new_active_version:
         # Handle deactivation of the previously active version
         await on_graph_deactivate(current_active_graph, user_id=user_id)
 
 
-async def _update_library_agent_version_and_settings(
-    user_id: str, agent_graph: graph_db.GraphModel
-) -> library_model.LibraryAgent:
-    library = await library_db.update_agent_version_in_library(
-        user_id, agent_graph.id, agent_graph.version
-    )
-    updated_settings = GraphSettings.from_graph(
-        graph=agent_graph,
-        hitl_safe_mode=library.settings.human_in_the_loop_safe_mode,
-        sensitive_action_safe_mode=library.settings.sensitive_action_safe_mode,
-    )
-    if updated_settings != library.settings:
-        library = await library_db.update_library_agent(
-            library_agent_id=library.id,
-            user_id=user_id,
-            settings=updated_settings,
-        )
-    return library
-
-
 @v1_router.patch(
     path="/graphs/{graph_id}/settings",
     summary="Update graph settings",

autogpt_platform/backend/backend/api/rest_api.py

@@ -40,6 +40,10 @@ import backend.data.user
 import backend.integrations.webhooks.utils
 import backend.util.service
 import backend.util.settings
+from backend.api.features.chat.completion_consumer import (
+    start_completion_consumer,
+    stop_completion_consumer,
+)
 from backend.blocks.llm import DEFAULT_LLM_MODEL
 from backend.data.model import Credentials
 from backend.integrations.providers import ProviderName
@@ -118,9 +122,21 @@ async def lifespan_context(app: fastapi.FastAPI):
     await backend.data.graph.migrate_llm_models(DEFAULT_LLM_MODEL)
     await backend.integrations.webhooks.utils.migrate_legacy_triggered_graphs()
 
+    # Start chat completion consumer for Redis Streams notifications
+    try:
+        await start_completion_consumer()
+    except Exception as e:
+        logger.warning(f"Could not start chat completion consumer: {e}")
+
     with launch_darkly_context():
         yield
 
+    # Stop chat completion consumer
+    try:
+        await stop_completion_consumer()
+    except Exception as e:
+        logger.warning(f"Error stopping chat completion consumer: {e}")
+
     try:
         await shutdown_cloud_storage_handler()
     except Exception as e:

autogpt_platform/backend/backend/api/ws_api.py

@@ -66,18 +66,24 @@ async def event_broadcaster(manager: ConnectionManager):
     execution_bus = AsyncRedisExecutionEventBus()
     notification_bus = AsyncRedisNotificationEventBus()
 
-    async def execution_worker():
+    try:
-        async for event in execution_bus.listen("*"):
-            await manager.send_execution_update(event)
 
-    async def notification_worker():
+        async def execution_worker():
-        async for notification in notification_bus.listen("*"):
+            async for event in execution_bus.listen("*"):
-            await manager.send_notification(
+                await manager.send_execution_update(event)
-                user_id=notification.user_id,
-                payload=notification.payload,
-            )
 
-    await asyncio.gather(execution_worker(), notification_worker())
+        async def notification_worker():
+            async for notification in notification_bus.listen("*"):
+                await manager.send_notification(
+                    user_id=notification.user_id,
+                    payload=notification.payload,
+                )
+
+        await asyncio.gather(execution_worker(), notification_worker())
+    finally:
+        # Ensure PubSub connections are closed on any exit to prevent leaks
+        await execution_bus.close()
+        await notification_bus.close()
 
 
 async def authenticate_websocket(websocket: WebSocket) -> str:

autogpt_platform/backend/backend/blocks/elevenlabs/_auth.py

@@ -0,0 +1,28 @@
+"""ElevenLabs integration blocks - test credentials and shared utilities."""
+
+from typing import Literal
+
+from pydantic import SecretStr
+
+from backend.data.model import APIKeyCredentials, CredentialsMetaInput
+from backend.integrations.providers import ProviderName
+
+TEST_CREDENTIALS = APIKeyCredentials(
+    id="01234567-89ab-cdef-0123-456789abcdef",
+    provider="elevenlabs",
+    api_key=SecretStr("mock-elevenlabs-api-key"),
+    title="Mock ElevenLabs API key",
+    expires_at=None,
+)
+
+TEST_CREDENTIALS_INPUT = {
+    "provider": TEST_CREDENTIALS.provider,
+    "id": TEST_CREDENTIALS.id,
+    "type": TEST_CREDENTIALS.type,
+    "title": TEST_CREDENTIALS.title,
+}
+
+ElevenLabsCredentials = APIKeyCredentials
+ElevenLabsCredentialsInput = CredentialsMetaInput[
+    Literal[ProviderName.ELEVENLABS], Literal["api_key"]
+]

autogpt_platform/backend/backend/blocks/encoder_block.py

@@ -0,0 +1,77 @@
+"""Text encoding block for converting special characters to escape sequences."""
+
+import codecs
+
+from backend.data.block import (
+    Block,
+    BlockCategory,
+    BlockOutput,
+    BlockSchemaInput,
+    BlockSchemaOutput,
+)
+from backend.data.model import SchemaField
+
+
+class TextEncoderBlock(Block):
+    """
+    Encodes a string by converting special characters into escape sequences.
+
+    This block is the inverse of TextDecoderBlock. It takes text containing
+    special characters (like newlines, tabs, etc.) and converts them into
+    their escape sequence representations (e.g., newline becomes \\n).
+    """
+
+    class Input(BlockSchemaInput):
+        """Input schema for TextEncoderBlock."""
+
+        text: str = SchemaField(
+            description="A string containing special characters to be encoded",
+            placeholder="Your text with newlines and quotes to encode",
+        )
+
+    class Output(BlockSchemaOutput):
+        """Output schema for TextEncoderBlock."""
+
+        encoded_text: str = SchemaField(
+            description="The encoded text with special characters converted to escape sequences"
+        )
+        error: str = SchemaField(description="Error message if encoding fails")
+
+    def __init__(self):
+        super().__init__(
+            id="5185f32e-4b65-4ecf-8fbb-873f003f09d6",
+            description="Encodes a string by converting special characters into escape sequences",
+            categories={BlockCategory.TEXT},
+            input_schema=TextEncoderBlock.Input,
+            output_schema=TextEncoderBlock.Output,
+            test_input={
+                "text": """Hello
+World!
+This is a "quoted" string."""
+            },
+            test_output=[
+                (
+                    "encoded_text",
+                    """Hello\\nWorld!\\nThis is a "quoted" string.""",
+                )
+            ],
+        )
+
+    async def run(self, input_data: Input, **kwargs) -> BlockOutput:
+        """
+        Encode the input text by converting special characters to escape sequences.
+
+        Args:
+            input_data: The input containing the text to encode.
+            **kwargs: Additional keyword arguments (unused).
+
+        Yields:
+            The encoded text with escape sequences, or an error message if encoding fails.
+        """
+        try:
+            encoded_text = codecs.encode(input_data.text, "unicode_escape").decode(
+                "utf-8"
+            )
+            yield "encoded_text", encoded_text
+        except Exception as e:
+            yield "error", f"Encoding error: {str(e)}"

autogpt_platform/backend/backend/blocks/linear/_api.py

@@ -162,8 +162,16 @@ class LinearClient:
                 "searchTerm": team_name,
             }
 
-            team_id = await self.query(query, variables)
+            result = await self.query(query, variables)
-            return team_id["teams"]["nodes"][0]["id"]
+            nodes = result["teams"]["nodes"]
+
+            if not nodes:
+                raise LinearAPIException(
+                    f"Team '{team_name}' not found. Check the team name or key and try again.",
+                    status_code=404,
+                )
+
+            return nodes[0]["id"]
         except LinearAPIException as e:
             raise e
 
@@ -240,17 +248,44 @@ class LinearClient:
         except LinearAPIException as e:
             raise e
 
-    async def try_search_issues(self, term: str) -> list[Issue]:
+    async def try_search_issues(
+        self,
+        term: str,
+        max_results: int = 10,
+        team_id: str | None = None,
+    ) -> list[Issue]:
         try:
             query = """
-                query SearchIssues($term: String!, $includeComments: Boolean!) {
+                query SearchIssues(
-                    searchIssues(term: $term, includeComments: $includeComments) {
+                    $term: String!,
+                    $first: Int,
+                    $teamId: String
+                ) {
+                    searchIssues(
+                        term: $term,
+                        first: $first,
+                        teamId: $teamId
+                    ) {
                         nodes {
                             id
                             identifier
                             title
                             description
                             priority
+                            createdAt
+                            state {
+                                id
+                                name
+                                type
+                            }
+                            project {
+                                id
+                                name
+                            }
+                            assignee {
+                                id
+                                name
+                            }
                         }
                     }
                 }
@@ -258,7 +293,8 @@ class LinearClient:
 
             variables: dict[str, Any] = {
                 "term": term,
-                "includeComments": True,
+                "first": max_results,
+                "teamId": team_id,
             }
 
             issues = await self.query(query, variables)

autogpt_platform/backend/backend/blocks/linear/issues.py

@@ -17,7 +17,7 @@ from ._config import (
     LinearScope,
     linear,
 )
-from .models import CreateIssueResponse, Issue
+from .models import CreateIssueResponse, Issue, State
 
 
 class LinearCreateIssueBlock(Block):
@@ -135,9 +135,20 @@ class LinearSearchIssuesBlock(Block):
             description="Linear credentials with read permissions",
             required_scopes={LinearScope.READ},
         )
+        max_results: int = SchemaField(
+            description="Maximum number of results to return",
+            default=10,
+            ge=1,
+            le=100,
+        )
+        team_name: str | None = SchemaField(
+            description="Optional team name to filter results (e.g., 'Internal', 'Open Source')",
+            default=None,
+        )
 
     class Output(BlockSchemaOutput):
         issues: list[Issue] = SchemaField(description="List of issues")
+        error: str = SchemaField(description="Error message if the search failed")
 
     def __init__(self):
         super().__init__(
@@ -145,8 +156,11 @@ class LinearSearchIssuesBlock(Block):
             description="Searches for issues on Linear",
             input_schema=self.Input,
             output_schema=self.Output,
+            categories={BlockCategory.PRODUCTIVITY, BlockCategory.ISSUE_TRACKING},
             test_input={
                 "term": "Test issue",
+                "max_results": 10,
+                "team_name": None,
                 "credentials": TEST_CREDENTIALS_INPUT_OAUTH,
             },
             test_credentials=TEST_CREDENTIALS_OAUTH,
@@ -156,10 +170,14 @@ class LinearSearchIssuesBlock(Block):
                     [
                         Issue(
                             id="abc123",
-                            identifier="abc123",
+                            identifier="TST-123",
                             title="Test issue",
                             description="Test description",
                             priority=1,
+                            state=State(
+                                id="state1", name="In Progress", type="started"
+                            ),
+                            createdAt="2026-01-15T10:00:00.000Z",
                         )
                     ],
                 )
@@ -168,10 +186,12 @@ class LinearSearchIssuesBlock(Block):
                 "search_issues": lambda *args, **kwargs: [
                     Issue(
                         id="abc123",
-                        identifier="abc123",
+                        identifier="TST-123",
                         title="Test issue",
                         description="Test description",
                         priority=1,
+                        state=State(id="state1", name="In Progress", type="started"),
+                        createdAt="2026-01-15T10:00:00.000Z",
                     )
                 ]
             },
@@ -181,10 +201,22 @@ class LinearSearchIssuesBlock(Block):
     async def search_issues(
         credentials: OAuth2Credentials | APIKeyCredentials,
         term: str,
+        max_results: int = 10,
+        team_name: str | None = None,
     ) -> list[Issue]:
         client = LinearClient(credentials=credentials)
-        response: list[Issue] = await client.try_search_issues(term=term)
+
-        return response
+        # Resolve team name to ID if provided
+        # Raises LinearAPIException with descriptive message if team not found
+        team_id: str | None = None
+        if team_name:
+            team_id = await client.try_get_team_by_name(team_name=team_name)
+
+        return await client.try_search_issues(
+            term=term,
+            max_results=max_results,
+            team_id=team_id,
+        )
 
     async def run(
         self,
@@ -196,7 +228,10 @@ class LinearSearchIssuesBlock(Block):
         """Execute the issue search"""
         try:
             issues = await self.search_issues(
-                credentials=credentials, term=input_data.term
+                credentials=credentials,
+                term=input_data.term,
+                max_results=input_data.max_results,
+                team_name=input_data.team_name,
             )
             yield "issues", issues
         except LinearAPIException as e:

autogpt_platform/backend/backend/blocks/linear/models.py

@@ -36,12 +36,21 @@ class Project(BaseModel):
     content: str | None = None
 
 
+class State(BaseModel):
+    id: str
+    name: str
+    type: str | None = (
+        None  # Workflow state type (e.g., "triage", "backlog", "started", "completed", "canceled")
+    )
+
+
 class Issue(BaseModel):
     id: str
     identifier: str
     title: str
     description: str | None
     priority: int
+    state: State | None = None
     project: Project | None = None
     createdAt: str | None = None
     comments: list[Comment] | None = None

autogpt_platform/backend/backend/blocks/llm.py

@@ -32,7 +32,7 @@ from backend.data.model import (
 from backend.integrations.providers import ProviderName
 from backend.util import json
 from backend.util.logging import TruncatedLogger
-from backend.util.prompt import compress_prompt, estimate_token_count
+from backend.util.prompt import compress_context, estimate_token_count
 from backend.util.text import TextFormatter
 
 logger = TruncatedLogger(logging.getLogger(__name__), "[LLM-Block]")
@@ -115,6 +115,7 @@ class LlmModel(str, Enum, metaclass=LlmModelMeta):
     CLAUDE_4_5_OPUS = "claude-opus-4-5-20251101"
     CLAUDE_4_5_SONNET = "claude-sonnet-4-5-20250929"
     CLAUDE_4_5_HAIKU = "claude-haiku-4-5-20251001"
+    CLAUDE_4_6_OPUS = "claude-opus-4-6"
     CLAUDE_3_HAIKU = "claude-3-haiku-20240307"
     # AI/ML API models
     AIML_API_QWEN2_5_72B = "Qwen/Qwen2.5-72B-Instruct-Turbo"
@@ -270,6 +271,9 @@ MODEL_METADATA = {
     LlmModel.CLAUDE_4_SONNET: ModelMetadata(
         "anthropic", 200000, 64000, "Claude Sonnet 4", "Anthropic", "Anthropic", 2
     ),  # claude-4-sonnet-20250514
+    LlmModel.CLAUDE_4_6_OPUS: ModelMetadata(
+        "anthropic", 200000, 128000, "Claude Opus 4.6", "Anthropic", "Anthropic", 3
+    ),  # claude-opus-4-6
     LlmModel.CLAUDE_4_5_OPUS: ModelMetadata(
         "anthropic", 200000, 64000, "Claude Opus 4.5", "Anthropic", "Anthropic", 3
     ),  # claude-opus-4-5-20251101
@@ -634,11 +638,18 @@ async def llm_call(
     context_window = llm_model.context_window
 
     if compress_prompt_to_fit:
-        prompt = compress_prompt(
+        result = await compress_context(
             messages=prompt,
             target_tokens=llm_model.context_window // 2,
-            lossy_ok=True,
+            client=None,  # Truncation-only, no LLM summarization
+            reserve=0,  # Caller handles response token budget separately
         )
+        if result.error:
+            logger.warning(
+                f"Prompt compression did not meet target: {result.error}. "
+                f"Proceeding with {result.token_count} tokens."
+            )
+        prompt = result.messages
 
     # Calculate available tokens based on context window and input length
     estimated_input_tokens = estimate_token_count(prompt)

autogpt_platform/backend/backend/blocks/media.py

@@ -1,246 +0,0 @@
-import os
-import tempfile
-from typing import Optional
-
-from moviepy.audio.io.AudioFileClip import AudioFileClip
-from moviepy.video.fx.Loop import Loop
-from moviepy.video.io.VideoFileClip import VideoFileClip
-
-from backend.data.block import (
-    Block,
-    BlockCategory,
-    BlockOutput,
-    BlockSchemaInput,
-    BlockSchemaOutput,
-)
-from backend.data.execution import ExecutionContext
-from backend.data.model import SchemaField
-from backend.util.file import MediaFileType, get_exec_file_path, store_media_file
-
-
-class MediaDurationBlock(Block):
-
-    class Input(BlockSchemaInput):
-        media_in: MediaFileType = SchemaField(
-            description="Media input (URL, data URI, or local path)."
-        )
-        is_video: bool = SchemaField(
-            description="Whether the media is a video (True) or audio (False).",
-            default=True,
-        )
-
-    class Output(BlockSchemaOutput):
-        duration: float = SchemaField(
-            description="Duration of the media file (in seconds)."
-        )
-
-    def __init__(self):
-        super().__init__(
-            id="d8b91fd4-da26-42d4-8ecb-8b196c6d84b6",
-            description="Block to get the duration of a media file.",
-            categories={BlockCategory.MULTIMEDIA},
-            input_schema=MediaDurationBlock.Input,
-            output_schema=MediaDurationBlock.Output,
-        )
-
-    async def run(
-        self,
-        input_data: Input,
-        *,
-        execution_context: ExecutionContext,
-        **kwargs,
-    ) -> BlockOutput:
-        # 1) Store the input media locally
-        local_media_path = await store_media_file(
-            file=input_data.media_in,
-            execution_context=execution_context,
-            return_format="for_local_processing",
-        )
-        assert execution_context.graph_exec_id is not None
-        media_abspath = get_exec_file_path(
-            execution_context.graph_exec_id, local_media_path
-        )
-
-        # 2) Load the clip
-        if input_data.is_video:
-            clip = VideoFileClip(media_abspath)
-        else:
-            clip = AudioFileClip(media_abspath)
-
-        yield "duration", clip.duration
-
-
-class LoopVideoBlock(Block):
-    """
-    Block for looping (repeating) a video clip until a given duration or number of loops.
-    """
-
-    class Input(BlockSchemaInput):
-        video_in: MediaFileType = SchemaField(
-            description="The input video (can be a URL, data URI, or local path)."
-        )
-        # Provide EITHER a `duration` or `n_loops` or both. We'll demonstrate `duration`.
-        duration: Optional[float] = SchemaField(
-            description="Target duration (in seconds) to loop the video to. If omitted, defaults to no looping.",
-            default=None,
-            ge=0.0,
-        )
-        n_loops: Optional[int] = SchemaField(
-            description="Number of times to repeat the video. If omitted, defaults to 1 (no repeat).",
-            default=None,
-            ge=1,
-        )
-
-    class Output(BlockSchemaOutput):
-        video_out: str = SchemaField(
-            description="Looped video returned either as a relative path or a data URI."
-        )
-
-    def __init__(self):
-        super().__init__(
-            id="8bf9eef6-5451-4213-b265-25306446e94b",
-            description="Block to loop a video to a given duration or number of repeats.",
-            categories={BlockCategory.MULTIMEDIA},
-            input_schema=LoopVideoBlock.Input,
-            output_schema=LoopVideoBlock.Output,
-        )
-
-    async def run(
-        self,
-        input_data: Input,
-        *,
-        execution_context: ExecutionContext,
-        **kwargs,
-    ) -> BlockOutput:
-        assert execution_context.graph_exec_id is not None
-        assert execution_context.node_exec_id is not None
-        graph_exec_id = execution_context.graph_exec_id
-        node_exec_id = execution_context.node_exec_id
-
-        # 1) Store the input video locally
-        local_video_path = await store_media_file(
-            file=input_data.video_in,
-            execution_context=execution_context,
-            return_format="for_local_processing",
-        )
-        input_abspath = get_exec_file_path(graph_exec_id, local_video_path)
-
-        # 2) Load the clip
-        clip = VideoFileClip(input_abspath)
-
-        # 3) Apply the loop effect
-        looped_clip = clip
-        if input_data.duration:
-            # Loop until we reach the specified duration
-            looped_clip = looped_clip.with_effects([Loop(duration=input_data.duration)])
-        elif input_data.n_loops:
-            looped_clip = looped_clip.with_effects([Loop(n=input_data.n_loops)])
-        else:
-            raise ValueError("Either 'duration' or 'n_loops' must be provided.")
-
-        assert isinstance(looped_clip, VideoFileClip)
-
-        # 4) Save the looped output
-        output_filename = MediaFileType(
-            f"{node_exec_id}_looped_{os.path.basename(local_video_path)}"
-        )
-        output_abspath = get_exec_file_path(graph_exec_id, output_filename)
-
-        looped_clip = looped_clip.with_audio(clip.audio)
-        looped_clip.write_videofile(output_abspath, codec="libx264", audio_codec="aac")
-
-        # Return output - for_block_output returns workspace:// if available, else data URI
-        video_out = await store_media_file(
-            file=output_filename,
-            execution_context=execution_context,
-            return_format="for_block_output",
-        )
-
-        yield "video_out", video_out
-
-
-class AddAudioToVideoBlock(Block):
-    """
-    Block that adds (attaches) an audio track to an existing video.
-    Optionally scale the volume of the new track.
-    """
-
-    class Input(BlockSchemaInput):
-        video_in: MediaFileType = SchemaField(
-            description="Video input (URL, data URI, or local path)."
-        )
-        audio_in: MediaFileType = SchemaField(
-            description="Audio input (URL, data URI, or local path)."
-        )
-        volume: float = SchemaField(
-            description="Volume scale for the newly attached audio track (1.0 = original).",
-            default=1.0,
-        )
-
-    class Output(BlockSchemaOutput):
-        video_out: MediaFileType = SchemaField(
-            description="Final video (with attached audio), as a path or data URI."
-        )
-
-    def __init__(self):
-        super().__init__(
-            id="3503748d-62b6-4425-91d6-725b064af509",
-            description="Block to attach an audio file to a video file using moviepy.",
-            categories={BlockCategory.MULTIMEDIA},
-            input_schema=AddAudioToVideoBlock.Input,
-            output_schema=AddAudioToVideoBlock.Output,
-        )
-
-    async def run(
-        self,
-        input_data: Input,
-        *,
-        execution_context: ExecutionContext,
-        **kwargs,
-    ) -> BlockOutput:
-        assert execution_context.graph_exec_id is not None
-        assert execution_context.node_exec_id is not None
-        graph_exec_id = execution_context.graph_exec_id
-        node_exec_id = execution_context.node_exec_id
-
-        # 1) Store the inputs locally
-        local_video_path = await store_media_file(
-            file=input_data.video_in,
-            execution_context=execution_context,
-            return_format="for_local_processing",
-        )
-        local_audio_path = await store_media_file(
-            file=input_data.audio_in,
-            execution_context=execution_context,
-            return_format="for_local_processing",
-        )
-
-        abs_temp_dir = os.path.join(tempfile.gettempdir(), "exec_file", graph_exec_id)
-        video_abspath = os.path.join(abs_temp_dir, local_video_path)
-        audio_abspath = os.path.join(abs_temp_dir, local_audio_path)
-
-        # 2) Load video + audio with moviepy
-        video_clip = VideoFileClip(video_abspath)
-        audio_clip = AudioFileClip(audio_abspath)
-        # Optionally scale volume
-        if input_data.volume != 1.0:
-            audio_clip = audio_clip.with_volume_scaled(input_data.volume)
-
-        # 3) Attach the new audio track
-        final_clip = video_clip.with_audio(audio_clip)
-
-        # 4) Write to output file
-        output_filename = MediaFileType(
-            f"{node_exec_id}_audio_attached_{os.path.basename(local_video_path)}"
-        )
-        output_abspath = os.path.join(abs_temp_dir, output_filename)
-        final_clip.write_videofile(output_abspath, codec="libx264", audio_codec="aac")
-
-        # 5) Return output - for_block_output returns workspace:// if available, else data URI
-        video_out = await store_media_file(
-            file=output_filename,
-            execution_context=execution_context,
-            return_format="for_block_output",
-        )
-
-        yield "video_out", video_out

autogpt_platform/backend/backend/blocks/stagehand/blocks.py

@@ -182,10 +182,7 @@ class StagehandObserveBlock(Block):
         **kwargs,
     ) -> BlockOutput:
 
-        logger.info(f"OBSERVE: Stagehand credentials: {stagehand_credentials}")
+        logger.debug(f"OBSERVE: Using model provider {model_credentials.provider}")
-        logger.info(
-            f"OBSERVE: Model credentials: {model_credentials} for provider {model_credentials.provider} secret: {model_credentials.api_key.get_secret_value()}"
-        )
 
         with disable_signal_handling():
             stagehand = Stagehand(
@@ -282,10 +279,7 @@ class StagehandActBlock(Block):
         **kwargs,
     ) -> BlockOutput:
 
-        logger.info(f"ACT: Stagehand credentials: {stagehand_credentials}")
+        logger.debug(f"ACT: Using model provider {model_credentials.provider}")
-        logger.info(
-            f"ACT: Model credentials: {model_credentials} for provider {model_credentials.provider} secret: {model_credentials.api_key.get_secret_value()}"
-        )
 
         with disable_signal_handling():
             stagehand = Stagehand(
@@ -370,10 +364,7 @@ class StagehandExtractBlock(Block):
         **kwargs,
     ) -> BlockOutput:
 
-        logger.info(f"EXTRACT: Stagehand credentials: {stagehand_credentials}")
+        logger.debug(f"EXTRACT: Using model provider {model_credentials.provider}")
-        logger.info(
-            f"EXTRACT: Model credentials: {model_credentials} for provider {model_credentials.provider} secret: {model_credentials.api_key.get_secret_value()}"
-        )
 
         with disable_signal_handling():
             stagehand = Stagehand(

autogpt_platform/backend/backend/blocks/test/test_text_encoder.py

@@ -0,0 +1,77 @@
+import pytest
+
+from backend.blocks.encoder_block import TextEncoderBlock
+
+
+@pytest.mark.asyncio
+async def test_text_encoder_basic():
+    """Test basic encoding of newlines and special characters."""
+    block = TextEncoderBlock()
+    result = []
+    async for output in block.run(TextEncoderBlock.Input(text="Hello\nWorld")):
+        result.append(output)
+
+    assert len(result) == 1
+    assert result[0][0] == "encoded_text"
+    assert result[0][1] == "Hello\\nWorld"
+
+
+@pytest.mark.asyncio
+async def test_text_encoder_multiple_escapes():
+    """Test encoding of multiple escape sequences."""
+    block = TextEncoderBlock()
+    result = []
+    async for output in block.run(
+        TextEncoderBlock.Input(text="Line1\nLine2\tTabbed\rCarriage")
+    ):
+        result.append(output)
+
+    assert len(result) == 1
+    assert result[0][0] == "encoded_text"
+    assert "\\n" in result[0][1]
+    assert "\\t" in result[0][1]
+    assert "\\r" in result[0][1]
+
+
+@pytest.mark.asyncio
+async def test_text_encoder_unicode():
+    """Test that unicode characters are handled correctly."""
+    block = TextEncoderBlock()
+    result = []
+    async for output in block.run(TextEncoderBlock.Input(text="Hello 世界\n")):
+        result.append(output)
+
+    assert len(result) == 1
+    assert result[0][0] == "encoded_text"
+    # Unicode characters should be escaped as \uXXXX sequences
+    assert "\\n" in result[0][1]
+
+
+@pytest.mark.asyncio
+async def test_text_encoder_empty_string():
+    """Test encoding of an empty string."""
+    block = TextEncoderBlock()
+    result = []
+    async for output in block.run(TextEncoderBlock.Input(text="")):
+        result.append(output)
+
+    assert len(result) == 1
+    assert result[0][0] == "encoded_text"
+    assert result[0][1] == ""
+
+
+@pytest.mark.asyncio
+async def test_text_encoder_error_handling():
+    """Test that encoding errors are handled gracefully."""
+    from unittest.mock import patch
+
+    block = TextEncoderBlock()
+    result = []
+
+    with patch("codecs.encode", side_effect=Exception("Mocked encoding error")):
+        async for output in block.run(TextEncoderBlock.Input(text="test")):
+            result.append(output)
+
+    assert len(result) == 1
+    assert result[0][0] == "error"
+    assert "Mocked encoding error" in result[0][1]

autogpt_platform/backend/backend/blocks/video/__init__.py

@@ -0,0 +1,37 @@
+"""Video editing blocks for AutoGPT Platform.
+
+This module provides blocks for:
+- Downloading videos from URLs (YouTube, Vimeo, news sites, direct links)
+- Clipping/trimming video segments
+- Concatenating multiple videos
+- Adding text overlays
+- Adding AI-generated narration
+- Getting media duration
+- Looping videos
+- Adding audio to videos
+
+Dependencies:
+- yt-dlp: For video downloading
+- moviepy: For video editing operations
+- elevenlabs: For AI narration (optional)
+"""
+
+from backend.blocks.video.add_audio import AddAudioToVideoBlock
+from backend.blocks.video.clip import VideoClipBlock
+from backend.blocks.video.concat import VideoConcatBlock
+from backend.blocks.video.download import VideoDownloadBlock
+from backend.blocks.video.duration import MediaDurationBlock
+from backend.blocks.video.loop import LoopVideoBlock
+from backend.blocks.video.narration import VideoNarrationBlock
+from backend.blocks.video.text_overlay import VideoTextOverlayBlock
+
+__all__ = [
+    "AddAudioToVideoBlock",
+    "LoopVideoBlock",
+    "MediaDurationBlock",
+    "VideoClipBlock",
+    "VideoConcatBlock",
+    "VideoDownloadBlock",
+    "VideoNarrationBlock",
+    "VideoTextOverlayBlock",
+]

autogpt_platform/backend/backend/blocks/video/_utils.py

@@ -0,0 +1,131 @@
+"""Shared utilities for video blocks."""
+
+from __future__ import annotations
+
+import logging
+import os
+import re
+import subprocess
+from pathlib import Path
+
+logger = logging.getLogger(__name__)
+
+# Known operation tags added by video blocks
+_VIDEO_OPS = (
+    r"(?:clip|overlay|narrated|looped|concat|audio_attached|with_audio|narration)"
+)
+
+# Matches: {node_exec_id}_{operation}_ where node_exec_id contains a UUID
+_BLOCK_PREFIX_RE = re.compile(
+    r"^[a-zA-Z0-9_-]*"
+    r"[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}"
+    r"[a-zA-Z0-9_-]*"
+    r"_" + _VIDEO_OPS + r"_"
+)
+
+# Matches: a lone {node_exec_id}_ prefix (no operation keyword, e.g. download output)
+_UUID_PREFIX_RE = re.compile(
+    r"^[a-zA-Z0-9_-]*"
+    r"[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}"
+    r"[a-zA-Z0-9_-]*_"
+)
+
+
+def extract_source_name(input_path: str, max_length: int = 50) -> str:
+    """Extract the original source filename by stripping block-generated prefixes.
+
+    Iteratively removes {node_exec_id}_{operation}_ prefixes that accumulate
+    when chaining video blocks, recovering the original human-readable name.
+
+    Safe for plain filenames (no UUID -> no stripping).
+    Falls back to "video" if everything is stripped.
+    """
+    stem = Path(input_path).stem
+
+    # Pass 1: strip {node_exec_id}_{operation}_ prefixes iteratively
+    while _BLOCK_PREFIX_RE.match(stem):
+        stem = _BLOCK_PREFIX_RE.sub("", stem, count=1)
+
+    # Pass 2: strip a lone {node_exec_id}_ prefix (e.g. from download block)
+    if _UUID_PREFIX_RE.match(stem):
+        stem = _UUID_PREFIX_RE.sub("", stem, count=1)
+
+    if not stem:
+        return "video"
+
+    return stem[:max_length]
+
+
+def get_video_codecs(output_path: str) -> tuple[str, str]:
+    """Get appropriate video and audio codecs based on output file extension.
+
+    Args:
+        output_path: Path to the output file (used to determine extension)
+
+    Returns:
+        Tuple of (video_codec, audio_codec)
+
+    Codec mappings:
+        - .mp4: H.264 + AAC (universal compatibility)
+        - .webm: VP8 + Vorbis (web streaming)
+        - .mkv: H.264 + AAC (container supports many codecs)
+        - .mov: H.264 + AAC (Apple QuickTime, widely compatible)
+        - .m4v: H.264 + AAC (Apple iTunes/devices)
+        - .avi: MPEG-4 + MP3 (legacy Windows)
+    """
+    ext = os.path.splitext(output_path)[1].lower()
+
+    codec_map: dict[str, tuple[str, str]] = {
+        ".mp4": ("libx264", "aac"),
+        ".webm": ("libvpx", "libvorbis"),
+        ".mkv": ("libx264", "aac"),
+        ".mov": ("libx264", "aac"),
+        ".m4v": ("libx264", "aac"),
+        ".avi": ("mpeg4", "libmp3lame"),
+    }
+
+    return codec_map.get(ext, ("libx264", "aac"))
+
+
+def strip_chapters_inplace(video_path: str) -> None:
+    """Strip chapter metadata from a media file in-place using ffmpeg.
+
+    MoviePy 2.x crashes with IndexError when parsing files with embedded
+    chapter metadata (https://github.com/Zulko/moviepy/issues/2419).
+    This strips chapters without re-encoding.
+
+    Args:
+        video_path: Absolute path to the media file to strip chapters from.
+    """
+    base, ext = os.path.splitext(video_path)
+    tmp_path = base + ".tmp" + ext
+    try:
+        result = subprocess.run(
+            [
+                "ffmpeg",
+                "-y",
+                "-i",
+                video_path,
+                "-map_chapters",
+                "-1",
+                "-codec",
+                "copy",
+                tmp_path,
+            ],
+            capture_output=True,
+            text=True,
+            timeout=300,
+        )
+        if result.returncode != 0:
+            logger.warning(
+                "ffmpeg chapter strip failed (rc=%d): %s",
+                result.returncode,
+                result.stderr,
+            )
+            return
+        os.replace(tmp_path, video_path)
+    except FileNotFoundError:
+        logger.warning("ffmpeg not found; skipping chapter strip")
+    finally:
+        if os.path.exists(tmp_path):
+            os.unlink(tmp_path)

autogpt_platform/backend/backend/blocks/video/add_audio.py

@@ -0,0 +1,113 @@
+"""AddAudioToVideoBlock - Attach an audio track to a video file."""
+
+from moviepy.audio.io.AudioFileClip import AudioFileClip
+from moviepy.video.io.VideoFileClip import VideoFileClip
+
+from backend.blocks.video._utils import extract_source_name, strip_chapters_inplace
+from backend.data.block import (
+    Block,
+    BlockCategory,
+    BlockOutput,
+    BlockSchemaInput,
+    BlockSchemaOutput,
+)
+from backend.data.execution import ExecutionContext
+from backend.data.model import SchemaField
+from backend.util.file import MediaFileType, get_exec_file_path, store_media_file
+
+
+class AddAudioToVideoBlock(Block):
+    """Add (attach) an audio track to an existing video."""
+
+    class Input(BlockSchemaInput):
+        video_in: MediaFileType = SchemaField(
+            description="Video input (URL, data URI, or local path)."
+        )
+        audio_in: MediaFileType = SchemaField(
+            description="Audio input (URL, data URI, or local path)."
+        )
+        volume: float = SchemaField(
+            description="Volume scale for the newly attached audio track (1.0 = original).",
+            default=1.0,
+        )
+
+    class Output(BlockSchemaOutput):
+        video_out: MediaFileType = SchemaField(
+            description="Final video (with attached audio), as a path or data URI."
+        )
+
+    def __init__(self):
+        super().__init__(
+            id="3503748d-62b6-4425-91d6-725b064af509",
+            description="Block to attach an audio file to a video file using moviepy.",
+            categories={BlockCategory.MULTIMEDIA},
+            input_schema=AddAudioToVideoBlock.Input,
+            output_schema=AddAudioToVideoBlock.Output,
+        )
+
+    async def run(
+        self,
+        input_data: Input,
+        *,
+        execution_context: ExecutionContext,
+        **kwargs,
+    ) -> BlockOutput:
+        assert execution_context.graph_exec_id is not None
+        assert execution_context.node_exec_id is not None
+        graph_exec_id = execution_context.graph_exec_id
+        node_exec_id = execution_context.node_exec_id
+
+        # 1) Store the inputs locally
+        local_video_path = await store_media_file(
+            file=input_data.video_in,
+            execution_context=execution_context,
+            return_format="for_local_processing",
+        )
+        local_audio_path = await store_media_file(
+            file=input_data.audio_in,
+            execution_context=execution_context,
+            return_format="for_local_processing",
+        )
+
+        video_abspath = get_exec_file_path(graph_exec_id, local_video_path)
+        audio_abspath = get_exec_file_path(graph_exec_id, local_audio_path)
+
+        # 2) Load video + audio with moviepy
+        strip_chapters_inplace(video_abspath)
+        strip_chapters_inplace(audio_abspath)
+        video_clip = None
+        audio_clip = None
+        final_clip = None
+        try:
+            video_clip = VideoFileClip(video_abspath)
+            audio_clip = AudioFileClip(audio_abspath)
+            # Optionally scale volume
+            if input_data.volume != 1.0:
+                audio_clip = audio_clip.with_volume_scaled(input_data.volume)
+
+            # 3) Attach the new audio track
+            final_clip = video_clip.with_audio(audio_clip)
+
+            # 4) Write to output file
+            source = extract_source_name(local_video_path)
+            output_filename = MediaFileType(f"{node_exec_id}_with_audio_{source}.mp4")
+            output_abspath = get_exec_file_path(graph_exec_id, output_filename)
+            final_clip.write_videofile(
+                output_abspath, codec="libx264", audio_codec="aac"
+            )
+        finally:
+            if final_clip:
+                final_clip.close()
+            if audio_clip:
+                audio_clip.close()
+            if video_clip:
+                video_clip.close()
+
+        # 5) Return output - for_block_output returns workspace:// if available, else data URI
+        video_out = await store_media_file(
+            file=output_filename,
+            execution_context=execution_context,
+            return_format="for_block_output",
+        )
+
+        yield "video_out", video_out

autogpt_platform/backend/backend/blocks/video/clip.py

@@ -0,0 +1,167 @@
+"""VideoClipBlock - Extract a segment from a video file."""
+
+from typing import Literal
+
+from moviepy.video.io.VideoFileClip import VideoFileClip
+
+from backend.blocks.video._utils import (
+    extract_source_name,
+    get_video_codecs,
+    strip_chapters_inplace,
+)
+from backend.data.block import (
+    Block,
+    BlockCategory,
+    BlockOutput,
+    BlockSchemaInput,
+    BlockSchemaOutput,
+)
+from backend.data.execution import ExecutionContext
+from backend.data.model import SchemaField
+from backend.util.exceptions import BlockExecutionError
+from backend.util.file import MediaFileType, get_exec_file_path, store_media_file
+
+
+class VideoClipBlock(Block):
+    """Extract a time segment from a video."""
+
+    class Input(BlockSchemaInput):
+        video_in: MediaFileType = SchemaField(
+            description="Input video (URL, data URI, or local path)"
+        )
+        start_time: float = SchemaField(description="Start time in seconds", ge=0.0)
+        end_time: float = SchemaField(description="End time in seconds", ge=0.0)
+        output_format: Literal["mp4", "webm", "mkv", "mov"] = SchemaField(
+            description="Output format", default="mp4", advanced=True
+        )
+
+    class Output(BlockSchemaOutput):
+        video_out: MediaFileType = SchemaField(
+            description="Clipped video file (path or data URI)"
+        )
+        duration: float = SchemaField(description="Clip duration in seconds")
+
+    def __init__(self):
+        super().__init__(
+            id="8f539119-e580-4d86-ad41-86fbcb22abb1",
+            description="Extract a time segment from a video",
+            categories={BlockCategory.MULTIMEDIA},
+            input_schema=self.Input,
+            output_schema=self.Output,
+            test_input={
+                "video_in": "/tmp/test.mp4",
+                "start_time": 0.0,
+                "end_time": 10.0,
+            },
+            test_output=[("video_out", str), ("duration", float)],
+            test_mock={
+                "_clip_video": lambda *args: 10.0,
+                "_store_input_video": lambda *args, **kwargs: "test.mp4",
+                "_store_output_video": lambda *args, **kwargs: "clip_test.mp4",
+            },
+        )
+
+    async def _store_input_video(
+        self, execution_context: ExecutionContext, file: MediaFileType
+    ) -> MediaFileType:
+        """Store input video. Extracted for testability."""
+        return await store_media_file(
+            file=file,
+            execution_context=execution_context,
+            return_format="for_local_processing",
+        )
+
+    async def _store_output_video(
+        self, execution_context: ExecutionContext, file: MediaFileType
+    ) -> MediaFileType:
+        """Store output video. Extracted for testability."""
+        return await store_media_file(
+            file=file,
+            execution_context=execution_context,
+            return_format="for_block_output",
+        )
+
+    def _clip_video(
+        self,
+        video_abspath: str,
+        output_abspath: str,
+        start_time: float,
+        end_time: float,
+    ) -> float:
+        """Extract a clip from a video. Extracted for testability."""
+        clip = None
+        subclip = None
+        try:
+            strip_chapters_inplace(video_abspath)
+            clip = VideoFileClip(video_abspath)
+            subclip = clip.subclipped(start_time, end_time)
+            video_codec, audio_codec = get_video_codecs(output_abspath)
+            subclip.write_videofile(
+                output_abspath, codec=video_codec, audio_codec=audio_codec
+            )
+            return subclip.duration
+        finally:
+            if subclip:
+                subclip.close()
+            if clip:
+                clip.close()
+
+    async def run(
+        self,
+        input_data: Input,
+        *,
+        execution_context: ExecutionContext,
+        node_exec_id: str,
+        **kwargs,
+    ) -> BlockOutput:
+        # Validate time range
+        if input_data.end_time <= input_data.start_time:
+            raise BlockExecutionError(
+                message=f"end_time ({input_data.end_time}) must be greater than start_time ({input_data.start_time})",
+                block_name=self.name,
+                block_id=str(self.id),
+            )
+
+        try:
+            assert execution_context.graph_exec_id is not None
+
+            # Store the input video locally
+            local_video_path = await self._store_input_video(
+                execution_context, input_data.video_in
+            )
+            video_abspath = get_exec_file_path(
+                execution_context.graph_exec_id, local_video_path
+            )
+
+            # Build output path
+            source = extract_source_name(local_video_path)
+            output_filename = MediaFileType(
+                f"{node_exec_id}_clip_{source}.{input_data.output_format}"
+            )
+            output_abspath = get_exec_file_path(
+                execution_context.graph_exec_id, output_filename
+            )
+
+            duration = self._clip_video(
+                video_abspath,
+                output_abspath,
+                input_data.start_time,
+                input_data.end_time,
+            )
+
+            # Return as workspace path or data URI based on context
+            video_out = await self._store_output_video(
+                execution_context, output_filename
+            )
+
+            yield "video_out", video_out
+            yield "duration", duration
+
+        except BlockExecutionError:
+            raise
+        except Exception as e:
+            raise BlockExecutionError(
+                message=f"Failed to clip video: {e}",
+                block_name=self.name,
+                block_id=str(self.id),
+            ) from e

autogpt_platform/backend/backend/blocks/video/concat.py

@@ -0,0 +1,227 @@
+"""VideoConcatBlock - Concatenate multiple video clips into one."""
+
+from typing import Literal
+
+from moviepy import concatenate_videoclips
+from moviepy.video.fx import CrossFadeIn, CrossFadeOut, FadeIn, FadeOut
+from moviepy.video.io.VideoFileClip import VideoFileClip
+
+from backend.blocks.video._utils import (
+    extract_source_name,
+    get_video_codecs,
+    strip_chapters_inplace,
+)
+from backend.data.block import (
+    Block,
+    BlockCategory,
+    BlockOutput,
+    BlockSchemaInput,
+    BlockSchemaOutput,
+)
+from backend.data.execution import ExecutionContext
+from backend.data.model import SchemaField
+from backend.util.exceptions import BlockExecutionError
+from backend.util.file import MediaFileType, get_exec_file_path, store_media_file
+
+
+class VideoConcatBlock(Block):
+    """Merge multiple video clips into one continuous video."""
+
+    class Input(BlockSchemaInput):
+        videos: list[MediaFileType] = SchemaField(
+            description="List of video files to concatenate (in order)"
+        )
+        transition: Literal["none", "crossfade", "fade_black"] = SchemaField(
+            description="Transition between clips", default="none"
+        )
+        transition_duration: int = SchemaField(
+            description="Transition duration in seconds",
+            default=1,
+            ge=0,
+            advanced=True,
+        )
+        output_format: Literal["mp4", "webm", "mkv", "mov"] = SchemaField(
+            description="Output format", default="mp4", advanced=True
+        )
+
+    class Output(BlockSchemaOutput):
+        video_out: MediaFileType = SchemaField(
+            description="Concatenated video file (path or data URI)"
+        )
+        total_duration: float = SchemaField(description="Total duration in seconds")
+
+    def __init__(self):
+        super().__init__(
+            id="9b0f531a-1118-487f-aeec-3fa63ea8900a",
+            description="Merge multiple video clips into one continuous video",
+            categories={BlockCategory.MULTIMEDIA},
+            input_schema=self.Input,
+            output_schema=self.Output,
+            test_input={
+                "videos": ["/tmp/a.mp4", "/tmp/b.mp4"],
+            },
+            test_output=[
+                ("video_out", str),
+                ("total_duration", float),
+            ],
+            test_mock={
+                "_concat_videos": lambda *args: 20.0,
+                "_store_input_video": lambda *args, **kwargs: "test.mp4",
+                "_store_output_video": lambda *args, **kwargs: "concat_test.mp4",
+            },
+        )
+
+    async def _store_input_video(
+        self, execution_context: ExecutionContext, file: MediaFileType
+    ) -> MediaFileType:
+        """Store input video. Extracted for testability."""
+        return await store_media_file(
+            file=file,
+            execution_context=execution_context,
+            return_format="for_local_processing",
+        )
+
+    async def _store_output_video(
+        self, execution_context: ExecutionContext, file: MediaFileType
+    ) -> MediaFileType:
+        """Store output video. Extracted for testability."""
+        return await store_media_file(
+            file=file,
+            execution_context=execution_context,
+            return_format="for_block_output",
+        )
+
+    def _concat_videos(
+        self,
+        video_abspaths: list[str],
+        output_abspath: str,
+        transition: str,
+        transition_duration: int,
+    ) -> float:
+        """Concatenate videos. Extracted for testability.
+
+        Returns:
+            Total duration of the concatenated video.
+        """
+        clips = []
+        faded_clips = []
+        final = None
+        try:
+            # Load clips
+            for v in video_abspaths:
+                strip_chapters_inplace(v)
+                clips.append(VideoFileClip(v))
+
+            # Validate transition_duration against shortest clip
+            if transition in {"crossfade", "fade_black"} and transition_duration > 0:
+                min_duration = min(c.duration for c in clips)
+                if transition_duration >= min_duration:
+                    raise BlockExecutionError(
+                        message=(
+                            f"transition_duration ({transition_duration}s) must be "
+                            f"shorter than the shortest clip ({min_duration:.2f}s)"
+                        ),
+                        block_name=self.name,
+                        block_id=str(self.id),
+                    )
+
+            if transition == "crossfade":
+                for i, clip in enumerate(clips):
+                    effects = []
+                    if i > 0:
+                        effects.append(CrossFadeIn(transition_duration))
+                    if i < len(clips) - 1:
+                        effects.append(CrossFadeOut(transition_duration))
+                    if effects:
+                        clip = clip.with_effects(effects)
+                    faded_clips.append(clip)
+                final = concatenate_videoclips(
+                    faded_clips,
+                    method="compose",
+                    padding=-transition_duration,
+                )
+            elif transition == "fade_black":
+                for clip in clips:
+                    faded = clip.with_effects(
+                        [FadeIn(transition_duration), FadeOut(transition_duration)]
+                    )
+                    faded_clips.append(faded)
+                final = concatenate_videoclips(faded_clips)
+            else:
+                final = concatenate_videoclips(clips)
+
+            video_codec, audio_codec = get_video_codecs(output_abspath)
+            final.write_videofile(
+                output_abspath, codec=video_codec, audio_codec=audio_codec
+            )
+
+            return final.duration
+        finally:
+            if final:
+                final.close()
+            for clip in faded_clips:
+                clip.close()
+            for clip in clips:
+                clip.close()
+
+    async def run(
+        self,
+        input_data: Input,
+        *,
+        execution_context: ExecutionContext,
+        node_exec_id: str,
+        **kwargs,
+    ) -> BlockOutput:
+        # Validate minimum clips
+        if len(input_data.videos) < 2:
+            raise BlockExecutionError(
+                message="At least 2 videos are required for concatenation",
+                block_name=self.name,
+                block_id=str(self.id),
+            )
+
+        try:
+            assert execution_context.graph_exec_id is not None
+
+            # Store all input videos locally
+            video_abspaths = []
+            for video in input_data.videos:
+                local_path = await self._store_input_video(execution_context, video)
+                video_abspaths.append(
+                    get_exec_file_path(execution_context.graph_exec_id, local_path)
+                )
+
+            # Build output path
+            source = (
+                extract_source_name(video_abspaths[0]) if video_abspaths else "video"
+            )
+            output_filename = MediaFileType(
+                f"{node_exec_id}_concat_{source}.{input_data.output_format}"
+            )
+            output_abspath = get_exec_file_path(
+                execution_context.graph_exec_id, output_filename
+            )
+
+            total_duration = self._concat_videos(
+                video_abspaths,
+                output_abspath,
+                input_data.transition,
+                input_data.transition_duration,
+            )
+
+            # Return as workspace path or data URI based on context
+            video_out = await self._store_output_video(
+                execution_context, output_filename
+            )
+
+            yield "video_out", video_out
+            yield "total_duration", total_duration
+
+        except BlockExecutionError:
+            raise
+        except Exception as e:
+            raise BlockExecutionError(
+                message=f"Failed to concatenate videos: {e}",
+                block_name=self.name,
+                block_id=str(self.id),
+            ) from e

autogpt_platform/backend/backend/blocks/video/download.py

@@ -0,0 +1,172 @@
+"""VideoDownloadBlock - Download video from URL (YouTube, Vimeo, news sites, direct links)."""
+
+import os
+import typing
+from typing import Literal
+
+import yt_dlp
+
+if typing.TYPE_CHECKING:
+    from yt_dlp import _Params
+
+from backend.data.block import (
+    Block,
+    BlockCategory,
+    BlockOutput,
+    BlockSchemaInput,
+    BlockSchemaOutput,
+)
+from backend.data.execution import ExecutionContext
+from backend.data.model import SchemaField
+from backend.util.exceptions import BlockExecutionError
+from backend.util.file import MediaFileType, get_exec_file_path, store_media_file
+
+
+class VideoDownloadBlock(Block):
+    """Download video from URL using yt-dlp."""
+
+    class Input(BlockSchemaInput):
+        url: str = SchemaField(
+            description="URL of the video to download (YouTube, Vimeo, direct link, etc.)",
+            placeholder="https://www.youtube.com/watch?v=...",
+        )
+        quality: Literal["best", "1080p", "720p", "480p", "audio_only"] = SchemaField(
+            description="Video quality preference", default="720p"
+        )
+        output_format: Literal["mp4", "webm", "mkv"] = SchemaField(
+            description="Output video format", default="mp4", advanced=True
+        )
+
+    class Output(BlockSchemaOutput):
+        video_file: MediaFileType = SchemaField(
+            description="Downloaded video (path or data URI)"
+        )
+        duration: float = SchemaField(description="Video duration in seconds")
+        title: str = SchemaField(description="Video title from source")
+        source_url: str = SchemaField(description="Original source URL")
+
+    def __init__(self):
+        super().__init__(
+            id="c35daabb-cd60-493b-b9ad-51f1fe4b50c4",
+            description="Download video from URL (YouTube, Vimeo, news sites, direct links)",
+            categories={BlockCategory.MULTIMEDIA},
+            input_schema=self.Input,
+            output_schema=self.Output,
+            disabled=True,  # Disable until we can sandbox yt-dlp and handle security implications
+            test_input={
+                "url": "https://www.youtube.com/watch?v=dQw4w9WgXcQ",
+                "quality": "480p",
+            },
+            test_output=[
+                ("video_file", str),
+                ("duration", float),
+                ("title", str),
+                ("source_url", str),
+            ],
+            test_mock={
+                "_download_video": lambda *args: (
+                    "video.mp4",
+                    212.0,
+                    "Test Video",
+                ),
+                "_store_output_video": lambda *args, **kwargs: "video.mp4",
+            },
+        )
+
+    async def _store_output_video(
+        self, execution_context: ExecutionContext, file: MediaFileType
+    ) -> MediaFileType:
+        """Store output video. Extracted for testability."""
+        return await store_media_file(
+            file=file,
+            execution_context=execution_context,
+            return_format="for_block_output",
+        )
+
+    def _get_format_string(self, quality: str) -> str:
+        formats = {
+            "best": "bestvideo+bestaudio/best",
+            "1080p": "bestvideo[height<=1080]+bestaudio/best[height<=1080]",
+            "720p": "bestvideo[height<=720]+bestaudio/best[height<=720]",
+            "480p": "bestvideo[height<=480]+bestaudio/best[height<=480]",
+            "audio_only": "bestaudio/best",
+        }
+        return formats.get(quality, formats["720p"])
+
+    def _download_video(
+        self,
+        url: str,
+        quality: str,
+        output_format: str,
+        output_dir: str,
+        node_exec_id: str,
+    ) -> tuple[str, float, str]:
+        """Download video. Extracted for testability."""
+        output_template = os.path.join(
+            output_dir, f"{node_exec_id}_%(title).50s.%(ext)s"
+        )
+
+        ydl_opts: "_Params" = {
+            "format": f"{self._get_format_string(quality)}/best",
+            "outtmpl": output_template,
+            "merge_output_format": output_format,
+            "quiet": True,
+            "no_warnings": True,
+        }
+
+        with yt_dlp.YoutubeDL(ydl_opts) as ydl:
+            info = ydl.extract_info(url, download=True)
+            video_path = ydl.prepare_filename(info)
+
+            # Handle format conversion in filename
+            if not video_path.endswith(f".{output_format}"):
+                video_path = video_path.rsplit(".", 1)[0] + f".{output_format}"
+
+            # Return just the filename, not the full path
+            filename = os.path.basename(video_path)
+
+            return (
+                filename,
+                info.get("duration") or 0.0,
+                info.get("title") or "Unknown",
+            )
+
+    async def run(
+        self,
+        input_data: Input,
+        *,
+        execution_context: ExecutionContext,
+        node_exec_id: str,
+        **kwargs,
+    ) -> BlockOutput:
+        try:
+            assert execution_context.graph_exec_id is not None
+
+            # Get the exec file directory
+            output_dir = get_exec_file_path(execution_context.graph_exec_id, "")
+            os.makedirs(output_dir, exist_ok=True)
+
+            filename, duration, title = self._download_video(
+                input_data.url,
+                input_data.quality,
+                input_data.output_format,
+                output_dir,
+                node_exec_id,
+            )
+
+            # Return as workspace path or data URI based on context
+            video_out = await self._store_output_video(
+                execution_context, MediaFileType(filename)
+            )
+
+            yield "video_file", video_out
+            yield "duration", duration
+            yield "title", title
+            yield "source_url", input_data.url
+
+        except Exception as e:
+            raise BlockExecutionError(
+                message=f"Failed to download video: {e}",
+                block_name=self.name,
+                block_id=str(self.id),
+            ) from e

autogpt_platform/backend/backend/blocks/video/duration.py

@@ -0,0 +1,77 @@
+"""MediaDurationBlock - Get the duration of a media file."""
+
+from moviepy.audio.io.AudioFileClip import AudioFileClip
+from moviepy.video.io.VideoFileClip import VideoFileClip
+
+from backend.blocks.video._utils import strip_chapters_inplace
+from backend.data.block import (
+    Block,
+    BlockCategory,
+    BlockOutput,
+    BlockSchemaInput,
+    BlockSchemaOutput,
+)
+from backend.data.execution import ExecutionContext
+from backend.data.model import SchemaField
+from backend.util.file import MediaFileType, get_exec_file_path, store_media_file
+
+
+class MediaDurationBlock(Block):
+    """Get the duration of a media file (video or audio)."""
+
+    class Input(BlockSchemaInput):
+        media_in: MediaFileType = SchemaField(
+            description="Media input (URL, data URI, or local path)."
+        )
+        is_video: bool = SchemaField(
+            description="Whether the media is a video (True) or audio (False).",
+            default=True,
+        )
+
+    class Output(BlockSchemaOutput):
+        duration: float = SchemaField(
+            description="Duration of the media file (in seconds)."
+        )
+
+    def __init__(self):
+        super().__init__(
+            id="d8b91fd4-da26-42d4-8ecb-8b196c6d84b6",
+            description="Block to get the duration of a media file.",
+            categories={BlockCategory.MULTIMEDIA},
+            input_schema=MediaDurationBlock.Input,
+            output_schema=MediaDurationBlock.Output,
+        )
+
+    async def run(
+        self,
+        input_data: Input,
+        *,
+        execution_context: ExecutionContext,
+        **kwargs,
+    ) -> BlockOutput:
+        # 1) Store the input media locally
+        local_media_path = await store_media_file(
+            file=input_data.media_in,
+            execution_context=execution_context,
+            return_format="for_local_processing",
+        )
+        assert execution_context.graph_exec_id is not None
+        media_abspath = get_exec_file_path(
+            execution_context.graph_exec_id, local_media_path
+        )
+
+        # 2) Strip chapters to avoid MoviePy crash, then load the clip
+        strip_chapters_inplace(media_abspath)
+        clip = None
+        try:
+            if input_data.is_video:
+                clip = VideoFileClip(media_abspath)
+            else:
+                clip = AudioFileClip(media_abspath)
+
+            duration = clip.duration
+        finally:
+            if clip:
+                clip.close()
+
+        yield "duration", duration

autogpt_platform/backend/backend/blocks/video/loop.py

@@ -0,0 +1,115 @@
+"""LoopVideoBlock - Loop a video to a given duration or number of repeats."""
+
+from typing import Optional
+
+from moviepy.video.fx.Loop import Loop
+from moviepy.video.io.VideoFileClip import VideoFileClip
+
+from backend.blocks.video._utils import extract_source_name, strip_chapters_inplace
+from backend.data.block import (
+    Block,
+    BlockCategory,
+    BlockOutput,
+    BlockSchemaInput,
+    BlockSchemaOutput,
+)
+from backend.data.execution import ExecutionContext
+from backend.data.model import SchemaField
+from backend.util.file import MediaFileType, get_exec_file_path, store_media_file
+
+
+class LoopVideoBlock(Block):
+    """Loop (repeat) a video clip until a given duration or number of loops."""
+
+    class Input(BlockSchemaInput):
+        video_in: MediaFileType = SchemaField(
+            description="The input video (can be a URL, data URI, or local path)."
+        )
+        duration: Optional[float] = SchemaField(
+            description="Target duration (in seconds) to loop the video to. Either duration or n_loops must be provided.",
+            default=None,
+            ge=0.0,
+            le=3600.0,  # Max 1 hour to prevent disk exhaustion
+        )
+        n_loops: Optional[int] = SchemaField(
+            description="Number of times to repeat the video. Either n_loops or duration must be provided.",
+            default=None,
+            ge=1,
+            le=10,  # Max 10 loops to prevent disk exhaustion
+        )
+
+    class Output(BlockSchemaOutput):
+        video_out: MediaFileType = SchemaField(
+            description="Looped video returned either as a relative path or a data URI."
+        )
+
+    def __init__(self):
+        super().__init__(
+            id="8bf9eef6-5451-4213-b265-25306446e94b",
+            description="Block to loop a video to a given duration or number of repeats.",
+            categories={BlockCategory.MULTIMEDIA},
+            input_schema=LoopVideoBlock.Input,
+            output_schema=LoopVideoBlock.Output,
+        )
+
+    async def run(
+        self,
+        input_data: Input,
+        *,
+        execution_context: ExecutionContext,
+        **kwargs,
+    ) -> BlockOutput:
+        assert execution_context.graph_exec_id is not None
+        assert execution_context.node_exec_id is not None
+        graph_exec_id = execution_context.graph_exec_id
+        node_exec_id = execution_context.node_exec_id
+
+        # 1) Store the input video locally
+        local_video_path = await store_media_file(
+            file=input_data.video_in,
+            execution_context=execution_context,
+            return_format="for_local_processing",
+        )
+        input_abspath = get_exec_file_path(graph_exec_id, local_video_path)
+
+        # 2) Load the clip
+        strip_chapters_inplace(input_abspath)
+        clip = None
+        looped_clip = None
+        try:
+            clip = VideoFileClip(input_abspath)
+
+            # 3) Apply the loop effect
+            if input_data.duration:
+                # Loop until we reach the specified duration
+                looped_clip = clip.with_effects([Loop(duration=input_data.duration)])
+            elif input_data.n_loops:
+                looped_clip = clip.with_effects([Loop(n=input_data.n_loops)])
+            else:
+                raise ValueError("Either 'duration' or 'n_loops' must be provided.")
+
+            assert isinstance(looped_clip, VideoFileClip)
+
+            # 4) Save the looped output
+            source = extract_source_name(local_video_path)
+            output_filename = MediaFileType(f"{node_exec_id}_looped_{source}.mp4")
+            output_abspath = get_exec_file_path(graph_exec_id, output_filename)
+
+            looped_clip = looped_clip.with_audio(clip.audio)
+            looped_clip.write_videofile(
+                output_abspath, codec="libx264", audio_codec="aac"
+            )
+        finally:
+            if looped_clip:
+                looped_clip.close()
+            if clip:
+                clip.close()
+
+        # Return output - for_block_output returns workspace:// if available, else data URI
+        video_out = await store_media_file(
+            file=output_filename,
+            execution_context=execution_context,
+            return_format="for_block_output",
+        )
+
+        yield "video_out", video_out

autogpt_platform/backend/backend/blocks/video/narration.py

@@ -0,0 +1,267 @@
+"""VideoNarrationBlock - Generate AI voice narration and add to video."""
+
+import os
+from typing import Literal
+
+from elevenlabs import ElevenLabs
+from moviepy import CompositeAudioClip
+from moviepy.audio.io.AudioFileClip import AudioFileClip
+from moviepy.video.io.VideoFileClip import VideoFileClip
+
+from backend.blocks.elevenlabs._auth import (
+    TEST_CREDENTIALS,
+    TEST_CREDENTIALS_INPUT,
+    ElevenLabsCredentials,
+    ElevenLabsCredentialsInput,
+)
+from backend.blocks.video._utils import (
+    extract_source_name,
+    get_video_codecs,
+    strip_chapters_inplace,
+)
+from backend.data.block import (
+    Block,
+    BlockCategory,
+    BlockOutput,
+    BlockSchemaInput,
+    BlockSchemaOutput,
+)
+from backend.data.execution import ExecutionContext
+from backend.data.model import CredentialsField, SchemaField
+from backend.util.exceptions import BlockExecutionError
+from backend.util.file import MediaFileType, get_exec_file_path, store_media_file
+
+
+class VideoNarrationBlock(Block):
+    """Generate AI narration and add to video."""
+
+    class Input(BlockSchemaInput):
+        credentials: ElevenLabsCredentialsInput = CredentialsField(
+            description="ElevenLabs API key for voice synthesis"
+        )
+        video_in: MediaFileType = SchemaField(
+            description="Input video (URL, data URI, or local path)"
+        )
+        script: str = SchemaField(description="Narration script text")
+        voice_id: str = SchemaField(
+            description="ElevenLabs voice ID", default="21m00Tcm4TlvDq8ikWAM"  # Rachel
+        )
+        model_id: Literal[
+            "eleven_multilingual_v2",
+            "eleven_flash_v2_5",
+            "eleven_turbo_v2_5",
+            "eleven_turbo_v2",
+        ] = SchemaField(
+            description="ElevenLabs TTS model",
+            default="eleven_multilingual_v2",
+        )
+        mix_mode: Literal["replace", "mix", "ducking"] = SchemaField(
+            description="How to combine with original audio. 'ducking' applies stronger attenuation than 'mix'.",
+            default="ducking",
+        )
+        narration_volume: float = SchemaField(
+            description="Narration volume (0.0 to 2.0)",
+            default=1.0,
+            ge=0.0,
+            le=2.0,
+            advanced=True,
+        )
+        original_volume: float = SchemaField(
+            description="Original audio volume when mixing (0.0 to 1.0)",
+            default=0.3,
+            ge=0.0,
+            le=1.0,
+            advanced=True,
+        )
+
+    class Output(BlockSchemaOutput):
+        video_out: MediaFileType = SchemaField(
+            description="Video with narration (path or data URI)"
+        )
+        audio_file: MediaFileType = SchemaField(
+            description="Generated audio file (path or data URI)"
+        )
+
+    def __init__(self):
+        super().__init__(
+            id="3d036b53-859c-4b17-9826-ca340f736e0e",
+            description="Generate AI narration and add to video",
+            categories={BlockCategory.MULTIMEDIA, BlockCategory.AI},
+            input_schema=self.Input,
+            output_schema=self.Output,
+            test_input={
+                "video_in": "/tmp/test.mp4",
+                "script": "Hello world",
+                "credentials": TEST_CREDENTIALS_INPUT,
+            },
+            test_credentials=TEST_CREDENTIALS,
+            test_output=[("video_out", str), ("audio_file", str)],
+            test_mock={
+                "_generate_narration_audio": lambda *args: b"mock audio content",
+                "_add_narration_to_video": lambda *args: None,
+                "_store_input_video": lambda *args, **kwargs: "test.mp4",
+                "_store_output_video": lambda *args, **kwargs: "narrated_test.mp4",
+            },
+        )
+
+    async def _store_input_video(
+        self, execution_context: ExecutionContext, file: MediaFileType
+    ) -> MediaFileType:
+        """Store input video. Extracted for testability."""
+        return await store_media_file(
+            file=file,
+            execution_context=execution_context,
+            return_format="for_local_processing",
+        )
+
+    async def _store_output_video(
+        self, execution_context: ExecutionContext, file: MediaFileType
+    ) -> MediaFileType:
+        """Store output video. Extracted for testability."""
+        return await store_media_file(
+            file=file,
+            execution_context=execution_context,
+            return_format="for_block_output",
+        )
+
+    def _generate_narration_audio(
+        self, api_key: str, script: str, voice_id: str, model_id: str
+    ) -> bytes:
+        """Generate narration audio via ElevenLabs API."""
+        client = ElevenLabs(api_key=api_key)
+        audio_generator = client.text_to_speech.convert(
+            voice_id=voice_id,
+            text=script,
+            model_id=model_id,
+        )
+        # The SDK returns a generator, collect all chunks
+        return b"".join(audio_generator)
+
+    def _add_narration_to_video(
+        self,
+        video_abspath: str,
+        audio_abspath: str,
+        output_abspath: str,
+        mix_mode: str,
+        narration_volume: float,
+        original_volume: float,
+    ) -> None:
+        """Add narration audio to video. Extracted for testability."""
+        video = None
+        final = None
+        narration_original = None
+        narration_scaled = None
+        original = None
+
+        try:
+            strip_chapters_inplace(video_abspath)
+            video = VideoFileClip(video_abspath)
+            narration_original = AudioFileClip(audio_abspath)
+            narration_scaled = narration_original.with_volume_scaled(narration_volume)
+            narration = narration_scaled
+
+            if mix_mode == "replace":
+                final_audio = narration
+            elif mix_mode == "mix":
+                if video.audio:
+                    original = video.audio.with_volume_scaled(original_volume)
+                    final_audio = CompositeAudioClip([original, narration])
+                else:
+                    final_audio = narration
+            else:  # ducking - apply stronger attenuation
+                if video.audio:
+                    # Ducking uses a much lower volume for original audio
+                    ducking_volume = original_volume * 0.3
+                    original = video.audio.with_volume_scaled(ducking_volume)
+                    final_audio = CompositeAudioClip([original, narration])
+                else:
+                    final_audio = narration
+
+            final = video.with_audio(final_audio)
+            video_codec, audio_codec = get_video_codecs(output_abspath)
+            final.write_videofile(
+                output_abspath, codec=video_codec, audio_codec=audio_codec
+            )
+
+        finally:
+            if original:
+                original.close()
+            if narration_scaled:
+                narration_scaled.close()
+            if narration_original:
+                narration_original.close()
+            if final:
+                final.close()
+            if video:
+                video.close()
+
+    async def run(
+        self,
+        input_data: Input,
+        *,
+        credentials: ElevenLabsCredentials,
+        execution_context: ExecutionContext,
+        node_exec_id: str,
+        **kwargs,
+    ) -> BlockOutput:
+        try:
+            assert execution_context.graph_exec_id is not None
+
+            # Store the input video locally
+            local_video_path = await self._store_input_video(
+                execution_context, input_data.video_in
+            )
+            video_abspath = get_exec_file_path(
+                execution_context.graph_exec_id, local_video_path
+            )
+
+            # Generate narration audio via ElevenLabs
+            audio_content = self._generate_narration_audio(
+                credentials.api_key.get_secret_value(),
+                input_data.script,
+                input_data.voice_id,
+                input_data.model_id,
+            )
+
+            # Save audio to exec file path
+            audio_filename = MediaFileType(f"{node_exec_id}_narration.mp3")
+            audio_abspath = get_exec_file_path(
+                execution_context.graph_exec_id, audio_filename
+            )
+            os.makedirs(os.path.dirname(audio_abspath), exist_ok=True)
+            with open(audio_abspath, "wb") as f:
+                f.write(audio_content)
+
+            # Add narration to video
+            source = extract_source_name(local_video_path)
+            output_filename = MediaFileType(f"{node_exec_id}_narrated_{source}.mp4")
+            output_abspath = get_exec_file_path(
+                execution_context.graph_exec_id, output_filename
+            )
+
+            self._add_narration_to_video(
+                video_abspath,
+                audio_abspath,
+                output_abspath,
+                input_data.mix_mode,
+                input_data.narration_volume,
+                input_data.original_volume,
+            )
+
+            # Return as workspace path or data URI based on context
+            video_out = await self._store_output_video(
+                execution_context, output_filename
+            )
+            audio_out = await self._store_output_video(
+                execution_context, audio_filename
+            )
+
+            yield "video_out", video_out
+            yield "audio_file", audio_out
+
+        except Exception as e:
+            raise BlockExecutionError(
+                message=f"Failed to add narration: {e}",
+                block_name=self.name,
+                block_id=str(self.id),
+            ) from e

autogpt_platform/backend/backend/blocks/video/text_overlay.py

@@ -0,0 +1,231 @@
+"""VideoTextOverlayBlock - Add text overlay to video."""
+
+from typing import Literal
+
+from moviepy import CompositeVideoClip, TextClip
+from moviepy.video.io.VideoFileClip import VideoFileClip
+
+from backend.blocks.video._utils import (
+    extract_source_name,
+    get_video_codecs,
+    strip_chapters_inplace,
+)
+from backend.data.block import (
+    Block,
+    BlockCategory,
+    BlockOutput,
+    BlockSchemaInput,
+    BlockSchemaOutput,
+)
+from backend.data.execution import ExecutionContext
+from backend.data.model import SchemaField
+from backend.util.exceptions import BlockExecutionError
+from backend.util.file import MediaFileType, get_exec_file_path, store_media_file
+
+
+class VideoTextOverlayBlock(Block):
+    """Add text overlay/caption to video."""
+
+    class Input(BlockSchemaInput):
+        video_in: MediaFileType = SchemaField(
+            description="Input video (URL, data URI, or local path)"
+        )
+        text: str = SchemaField(description="Text to overlay on video")
+        position: Literal[
+            "top",
+            "center",
+            "bottom",
+            "top-left",
+            "top-right",
+            "bottom-left",
+            "bottom-right",
+        ] = SchemaField(description="Position of text on screen", default="bottom")
+        start_time: float | None = SchemaField(
+            description="When to show text (seconds). None = entire video",
+            default=None,
+            advanced=True,
+        )
+        end_time: float | None = SchemaField(
+            description="When to hide text (seconds). None = until end",
+            default=None,
+            advanced=True,
+        )
+        font_size: int = SchemaField(
+            description="Font size", default=48, ge=12, le=200, advanced=True
+        )
+        font_color: str = SchemaField(
+            description="Font color (hex or name)", default="white", advanced=True
+        )
+        bg_color: str | None = SchemaField(
+            description="Background color behind text (None for transparent)",
+            default=None,
+            advanced=True,
+        )
+
+    class Output(BlockSchemaOutput):
+        video_out: MediaFileType = SchemaField(
+            description="Video with text overlay (path or data URI)"
+        )
+
+    def __init__(self):
+        super().__init__(
+            id="8ef14de6-cc90-430a-8cfa-3a003be92454",
+            description="Add text overlay/caption to video",
+            categories={BlockCategory.MULTIMEDIA},
+            input_schema=self.Input,
+            output_schema=self.Output,
+            disabled=True,  # Disable until we can lockdown imagemagick security policy
+            test_input={"video_in": "/tmp/test.mp4", "text": "Hello World"},
+            test_output=[("video_out", str)],
+            test_mock={
+                "_add_text_overlay": lambda *args: None,
+                "_store_input_video": lambda *args, **kwargs: "test.mp4",
+                "_store_output_video": lambda *args, **kwargs: "overlay_test.mp4",
+            },
+        )
+
+    async def _store_input_video(
+        self, execution_context: ExecutionContext, file: MediaFileType
+    ) -> MediaFileType:
+        """Store input video. Extracted for testability."""
+        return await store_media_file(
+            file=file,
+            execution_context=execution_context,
+            return_format="for_local_processing",
+        )
+
+    async def _store_output_video(
+        self, execution_context: ExecutionContext, file: MediaFileType
+    ) -> MediaFileType:
+        """Store output video. Extracted for testability."""
+        return await store_media_file(
+            file=file,
+            execution_context=execution_context,
+            return_format="for_block_output",
+        )
+
+    def _add_text_overlay(
+        self,
+        video_abspath: str,
+        output_abspath: str,
+        text: str,
+        position: str,
+        start_time: float | None,
+        end_time: float | None,
+        font_size: int,
+        font_color: str,
+        bg_color: str | None,
+    ) -> None:
+        """Add text overlay to video. Extracted for testability."""
+        video = None
+        final = None
+        txt_clip = None
+        try:
+            strip_chapters_inplace(video_abspath)
+            video = VideoFileClip(video_abspath)
+
+            txt_clip = TextClip(
+                text=text,
+                font_size=font_size,
+                color=font_color,
+                bg_color=bg_color,
+            )
+
+            # Position mapping
+            pos_map = {
+                "top": ("center", "top"),
+                "center": ("center", "center"),
+                "bottom": ("center", "bottom"),
+                "top-left": ("left", "top"),
+                "top-right": ("right", "top"),
+                "bottom-left": ("left", "bottom"),
+                "bottom-right": ("right", "bottom"),
+            }
+
+            txt_clip = txt_clip.with_position(pos_map[position])
+
+            # Set timing
+            start = start_time or 0
+            end = end_time or video.duration
+            duration = max(0, end - start)
+            txt_clip = txt_clip.with_start(start).with_end(end).with_duration(duration)
+
+            final = CompositeVideoClip([video, txt_clip])
+            video_codec, audio_codec = get_video_codecs(output_abspath)
+            final.write_videofile(
+                output_abspath, codec=video_codec, audio_codec=audio_codec
+            )
+
+        finally:
+            if txt_clip:
+                txt_clip.close()
+            if final:
+                final.close()
+            if video:
+                video.close()
+
+    async def run(
+        self,
+        input_data: Input,
+        *,
+        execution_context: ExecutionContext,
+        node_exec_id: str,
+        **kwargs,
+    ) -> BlockOutput:
+        # Validate time range if both are provided
+        if (
+            input_data.start_time is not None
+            and input_data.end_time is not None
+            and input_data.end_time <= input_data.start_time
+        ):
+            raise BlockExecutionError(
+                message=f"end_time ({input_data.end_time}) must be greater than start_time ({input_data.start_time})",
+                block_name=self.name,
+                block_id=str(self.id),
+            )
+
+        try:
+            assert execution_context.graph_exec_id is not None
+
+            # Store the input video locally
+            local_video_path = await self._store_input_video(
+                execution_context, input_data.video_in
+            )
+            video_abspath = get_exec_file_path(
+                execution_context.graph_exec_id, local_video_path
+            )
+
+            # Build output path
+            source = extract_source_name(local_video_path)
+            output_filename = MediaFileType(f"{node_exec_id}_overlay_{source}.mp4")
+            output_abspath = get_exec_file_path(
+                execution_context.graph_exec_id, output_filename
+            )
+
+            self._add_text_overlay(
+                video_abspath,
+                output_abspath,
+                input_data.text,
+                input_data.position,
+                input_data.start_time,
+                input_data.end_time,
+                input_data.font_size,
+                input_data.font_color,
+                input_data.bg_color,
+            )
+
+            # Return as workspace path or data URI based on context
+            video_out = await self._store_output_video(
+                execution_context, output_filename
+            )
+
+            yield "video_out", video_out
+
+        except BlockExecutionError:
+            raise
+        except Exception as e:
+            raise BlockExecutionError(
+                message=f"Failed to add text overlay: {e}",
+                block_name=self.name,
+                block_id=str(self.id),
+            ) from e

autogpt_platform/backend/backend/blocks/youtube.py

@@ -165,10 +165,13 @@ class TranscribeYoutubeVideoBlock(Block):
         credentials: WebshareProxyCredentials,
         **kwargs,
     ) -> BlockOutput:
-        video_id = self.extract_video_id(input_data.youtube_url)
+        try:
-        yield "video_id", video_id
+            video_id = self.extract_video_id(input_data.youtube_url)
+            transcript = self.get_transcript(video_id, credentials)
+            transcript_text = self.format_transcript(transcript=transcript)
 
-        transcript = self.get_transcript(video_id, credentials)
+            # Only yield after all operations succeed
-        transcript_text = self.format_transcript(transcript=transcript)
+            yield "video_id", video_id
-
+            yield "transcript", transcript_text
-        yield "transcript", transcript_text
+        except Exception as e:
+            yield "error", str(e)

autogpt_platform/backend/backend/data/block.py

@@ -873,14 +873,13 @@ def is_block_auth_configured(
 
 
 async def initialize_blocks() -> None:
-    # First, sync all provider costs to blocks
-    # Imported here to avoid circular import
     from backend.sdk.cost_integration import sync_all_provider_costs
+    from backend.util.retry import func_retry
 
     sync_all_provider_costs()
 
-    for cls in get_blocks().values():
+    @func_retry
-        block = cls()
+    async def sync_block_to_db(block: Block) -> None:
         existing_block = await AgentBlock.prisma().find_first(
             where={"OR": [{"id": block.id}, {"name": block.name}]}
         )
@@ -893,7 +892,7 @@ async def initialize_blocks() -> None:
                     outputSchema=json.dumps(block.output_schema.jsonschema()),
                 )
             )
-            continue
+            return
 
         input_schema = json.dumps(block.input_schema.jsonschema())
         output_schema = json.dumps(block.output_schema.jsonschema())
@@ -913,6 +912,25 @@ async def initialize_blocks() -> None:
                 },
             )
 
+    failed_blocks: list[str] = []
+    for cls in get_blocks().values():
+        block = cls()
+        try:
+            await sync_block_to_db(block)
+        except Exception as e:
+            logger.warning(
+                f"Failed to sync block {block.name} to database: {e}. "
+                "Block is still available in memory.",
+                exc_info=True,
+            )
+            failed_blocks.append(block.name)
+
+    if failed_blocks:
+        logger.error(
+            f"Failed to sync {len(failed_blocks)} block(s) to database: "
+            f"{', '.join(failed_blocks)}. These blocks are still available in memory."
+        )
+
 
 # Note on the return type annotation: https://github.com/microsoft/pyright/issues/10281
 def get_block(block_id: str) -> AnyBlockSchema | None:

autogpt_platform/backend/backend/data/block_cost_config.py

@@ -36,12 +36,14 @@ from backend.blocks.replicate.replicate_block import ReplicateModelBlock
 from backend.blocks.smart_decision_maker import SmartDecisionMakerBlock
 from backend.blocks.talking_head import CreateTalkingAvatarVideoBlock
 from backend.blocks.text_to_speech_block import UnrealTextToSpeechBlock
+from backend.blocks.video.narration import VideoNarrationBlock
 from backend.data.block import Block, BlockCost, BlockCostType
 from backend.integrations.credentials_store import (
     aiml_api_credentials,
     anthropic_credentials,
     apollo_credentials,
     did_credentials,
+    elevenlabs_credentials,
     enrichlayer_credentials,
     groq_credentials,
     ideogram_credentials,
@@ -78,6 +80,7 @@ MODEL_COST: dict[LlmModel, int] = {
     LlmModel.CLAUDE_4_1_OPUS: 21,
     LlmModel.CLAUDE_4_OPUS: 21,
     LlmModel.CLAUDE_4_SONNET: 5,
+    LlmModel.CLAUDE_4_6_OPUS: 14,
     LlmModel.CLAUDE_4_5_HAIKU: 4,
     LlmModel.CLAUDE_4_5_OPUS: 14,
     LlmModel.CLAUDE_4_5_SONNET: 9,
@@ -639,4 +642,16 @@ BLOCK_COSTS: dict[Type[Block], list[BlockCost]] = {
             },
         ),
     ],
+    VideoNarrationBlock: [
+        BlockCost(
+            cost_amount=5,  # ElevenLabs TTS cost
+            cost_filter={
+                "credentials": {
+                    "id": elevenlabs_credentials.id,
+                    "provider": elevenlabs_credentials.provider,
+                    "type": elevenlabs_credentials.type,
+                }
+            },
+        )
+    ],
 }

autogpt_platform/backend/backend/data/credit_test.py

@@ -134,6 +134,16 @@ async def test_block_credit_reset(server: SpinTestServer):
         month1 = datetime.now(timezone.utc).replace(month=1, day=1)
         user_credit.time_now = lambda: month1
 
+        # IMPORTANT: Set updatedAt to December of previous year to ensure it's
+        # in a different month than month1 (January). This fixes a timing bug
+        # where if the test runs in early February, 35 days ago would be January,
+        # matching the mocked month1 and preventing the refill from triggering.
+        dec_previous_year = month1.replace(year=month1.year - 1, month=12, day=15)
+        await UserBalance.prisma().update(
+            where={"userId": DEFAULT_USER_ID},
+            data={"updatedAt": dec_previous_year},
+        )
+
         # First call in month 1 should trigger refill
         balance = await user_credit.get_credits(DEFAULT_USER_ID)
         assert balance == REFILL_VALUE  # Should get 1000 credits

autogpt_platform/backend/backend/data/event_bus.py

@@ -133,10 +133,23 @@ class RedisEventBus(BaseRedisEventBus[M], ABC):
 
 
 class AsyncRedisEventBus(BaseRedisEventBus[M], ABC):
+    def __init__(self):
+        self._pubsub: AsyncPubSub | None = None
+
     @property
     async def connection(self) -> redis.AsyncRedis:
         return await redis.get_redis_async()
 
+    async def close(self) -> None:
+        """Close the PubSub connection if it exists."""
+        if self._pubsub is not None:
+            try:
+                await self._pubsub.close()
+            except Exception:
+                logger.warning("Failed to close PubSub connection", exc_info=True)
+            finally:
+                self._pubsub = None
+
     async def publish_event(self, event: M, channel_key: str):
         """
         Publish an event to Redis. Gracefully handles connection failures
@@ -157,6 +170,7 @@ class AsyncRedisEventBus(BaseRedisEventBus[M], ABC):
             await self.connection, channel_key
         )
         assert isinstance(pubsub, AsyncPubSub)
+        self._pubsub = pubsub
 
         if "*" in channel_key:
             await pubsub.psubscribe(full_channel_name)

autogpt_platform/backend/backend/data/graph.py

@@ -1028,6 +1028,39 @@ async def get_graph(
     return GraphModel.from_db(graph, for_export)
 
 
+async def get_store_listed_graphs(*graph_ids: str) -> dict[str, GraphModel]:
+    """Batch-fetch multiple store-listed graphs by their IDs.
+
+    Only returns graphs that have approved store listings (publicly available).
+    Does not require permission checks since store-listed graphs are public.
+
+    Args:
+        *graph_ids: Variable number of graph IDs to fetch
+
+    Returns:
+        Dict mapping graph_id to GraphModel for graphs with approved store listings
+    """
+    if not graph_ids:
+        return {}
+
+    store_listings = await StoreListingVersion.prisma().find_many(
+        where={
+            "agentGraphId": {"in": list(graph_ids)},
+            "submissionStatus": SubmissionStatus.APPROVED,
+            "isDeleted": False,
+        },
+        include={"AgentGraph": {"include": AGENT_GRAPH_INCLUDE}},
+        distinct=["agentGraphId"],
+        order={"agentGraphVersion": "desc"},
+    )
+
+    return {
+        listing.agentGraphId: GraphModel.from_db(listing.AgentGraph)
+        for listing in store_listings
+        if listing.AgentGraph
+    }
+
+
 async def get_graph_as_admin(
     graph_id: str,
     version: int | None = None,

autogpt_platform/backend/backend/data/model.py

@@ -19,7 +19,6 @@ from typing import (
     cast,
     get_args,
 )
-from urllib.parse import urlparse
 from uuid import uuid4
 
 from prisma.enums import CreditTransactionType, OnboardingStep
@@ -42,6 +41,7 @@ from typing_extensions import TypedDict
 
 from backend.integrations.providers import ProviderName
 from backend.util.json import loads as json_loads
+from backend.util.request import parse_url
 from backend.util.settings import Secrets
 
 # Type alias for any provider name (including custom ones)
@@ -397,19 +397,25 @@ class HostScopedCredentials(_BaseCredentials):
     def matches_url(self, url: str) -> bool:
         """Check if this credential should be applied to the given URL."""
 
-        parsed_url = urlparse(url)
+        request_host, request_port = _extract_host_from_url(url)
-        # Extract hostname without port
+        cred_scope_host, cred_scope_port = _extract_host_from_url(self.host)
-        request_host = parsed_url.hostname
         if not request_host:
             return False
 
-        # Simple host matching - exact match or wildcard subdomain match
+        # If a port is specified in credential host, the request host port must match
-        if self.host == request_host:
+        if cred_scope_port is not None and request_port != cred_scope_port:
+            return False
+        # Non-standard ports are only allowed if explicitly specified in credential host
+        elif cred_scope_port is None and request_port not in (80, 443, None):
+            return False
+
+        # Simple host matching
+        if cred_scope_host == request_host:
             return True
 
         # Support wildcard matching (e.g., "*.example.com" matches "api.example.com")
-        if self.host.startswith("*."):
+        if cred_scope_host.startswith("*."):
-            domain = self.host[2:]  # Remove "*."
+            domain = cred_scope_host[2:]  # Remove "*."
             return request_host.endswith(f".{domain}") or request_host == domain
 
         return False
@@ -551,13 +557,13 @@ class CredentialsMetaInput(BaseModel, Generic[CP, CT]):
     )
 
 
-def _extract_host_from_url(url: str) -> str:
+def _extract_host_from_url(url: str) -> tuple[str, int | None]:
-    """Extract host from URL for grouping host-scoped credentials."""
+    """Extract host and port from URL for grouping host-scoped credentials."""
     try:
-        parsed = urlparse(url)
+        parsed = parse_url(url)
-        return parsed.hostname or url
+        return parsed.hostname or url, parsed.port
     except Exception:
-        return ""
+        return "", None
 
 
 class CredentialsFieldInfo(BaseModel, Generic[CP, CT]):
@@ -606,7 +612,7 @@ class CredentialsFieldInfo(BaseModel, Generic[CP, CT]):
                 providers = frozenset(
                     [cast(CP, "http")]
                     + [
-                        cast(CP, _extract_host_from_url(str(value)))
+                        cast(CP, parse_url(str(value)).netloc)
                         for value in field.discriminator_values
                     ]
                 )

autogpt_platform/backend/backend/data/model_test.py

@@ -79,10 +79,23 @@ class TestHostScopedCredentials:
             headers={"Authorization": SecretStr("Bearer token")},
         )
 
-        assert creds.matches_url("http://localhost:8080/api/v1")
+        # Non-standard ports require explicit port in credential host
+        assert not creds.matches_url("http://localhost:8080/api/v1")
         assert creds.matches_url("https://localhost:443/secure/endpoint")
         assert creds.matches_url("http://localhost/simple")
 
+    def test_matches_url_with_explicit_port(self):
+        """Test URL matching with explicit port in credential host."""
+        creds = HostScopedCredentials(
+            provider="custom",
+            host="localhost:8080",
+            headers={"Authorization": SecretStr("Bearer token")},
+        )
+
+        assert creds.matches_url("http://localhost:8080/api/v1")
+        assert not creds.matches_url("http://localhost:3000/api/v1")
+        assert not creds.matches_url("http://localhost/simple")
+
     def test_empty_headers_dict(self):
         """Test HostScopedCredentials with empty headers."""
         creds = HostScopedCredentials(
@@ -128,8 +141,20 @@ class TestHostScopedCredentials:
             ("*.example.com", "https://sub.api.example.com/test", True),
             ("*.example.com", "https://example.com/test", True),
             ("*.example.com", "https://example.org/test", False),
-            ("localhost", "http://localhost:3000/test", True),
+            # Non-standard ports require explicit port in credential host
+            ("localhost", "http://localhost:3000/test", False),
+            ("localhost:3000", "http://localhost:3000/test", True),
             ("localhost", "http://127.0.0.1:3000/test", False),
+            # IPv6 addresses (frontend stores with brackets via URL.hostname)
+            ("[::1]", "http://[::1]/test", True),
+            ("[::1]", "http://[::1]:80/test", True),
+            ("[::1]", "https://[::1]:443/test", True),
+            ("[::1]", "http://[::1]:8080/test", False),  # Non-standard port
+            ("[::1]:8080", "http://[::1]:8080/test", True),
+            ("[::1]:8080", "http://[::1]:9090/test", False),
+            ("[2001:db8::1]", "http://[2001:db8::1]/path", True),
+            ("[2001:db8::1]", "https://[2001:db8::1]:443/path", True),
+            ("[2001:db8::1]", "http://[2001:db8::ff]/path", False),
         ],
     )
     def test_url_matching_parametrized(self, host: str, test_url: str, expected: bool):

autogpt_platform/backend/backend/executor/database.py

@@ -17,6 +17,7 @@ from backend.data.analytics import (
     get_accuracy_trends_and_alerts,
     get_marketplace_graphs_for_monitoring,
 )
+from backend.data.auth.oauth import cleanup_expired_oauth_tokens
 from backend.data.credit import UsageTransactionMetadata, get_user_credit_model
 from backend.data.execution import (
     create_graph_execution,
@@ -219,6 +220,9 @@ class DatabaseManager(AppService):
     # Onboarding
     increment_onboarding_runs = _(increment_onboarding_runs)
 
+    # OAuth
+    cleanup_expired_oauth_tokens = _(cleanup_expired_oauth_tokens)
+
     # Store
     get_store_agents = _(get_store_agents)
     get_store_agent_details = _(get_store_agent_details)
@@ -349,6 +353,9 @@ class DatabaseManagerAsyncClient(AppServiceClient):
     # Onboarding
     increment_onboarding_runs = d.increment_onboarding_runs
 
+    # OAuth
+    cleanup_expired_oauth_tokens = d.cleanup_expired_oauth_tokens
+
     # Store
     get_store_agents = d.get_store_agents
     get_store_agent_details = d.get_store_agent_details

autogpt_platform/backend/backend/executor/scheduler.py

@@ -24,11 +24,9 @@ from dotenv import load_dotenv
 from pydantic import BaseModel, Field, ValidationError
 from sqlalchemy import MetaData, create_engine
 
-from backend.data.auth.oauth import cleanup_expired_oauth_tokens
 from backend.data.block import BlockInput
 from backend.data.execution import GraphExecutionWithNodes
 from backend.data.model import CredentialsMetaInput
-from backend.data.onboarding import increment_onboarding_runs
 from backend.executor import utils as execution_utils
 from backend.monitoring import (
     NotificationJobArgs,
@@ -38,7 +36,11 @@ from backend.monitoring import (
     report_execution_accuracy_alerts,
     report_late_executions,
 )
-from backend.util.clients import get_database_manager_client, get_scheduler_client
+from backend.util.clients import (
+    get_database_manager_async_client,
+    get_database_manager_client,
+    get_scheduler_client,
+)
 from backend.util.cloud_storage import cleanup_expired_files_async
 from backend.util.exceptions import (
     GraphNotFoundError,
@@ -148,6 +150,7 @@ def execute_graph(**kwargs):
 async def _execute_graph(**kwargs):
     args = GraphExecutionJobArgs(**kwargs)
     start_time = asyncio.get_event_loop().time()
+    db = get_database_manager_async_client()
     try:
         logger.info(f"Executing recurring job for graph #{args.graph_id}")
         graph_exec: GraphExecutionWithNodes = await execution_utils.add_graph_execution(
@@ -157,7 +160,7 @@ async def _execute_graph(**kwargs):
             inputs=args.input_data,
             graph_credentials_inputs=args.input_credentials,
         )
-        await increment_onboarding_runs(args.user_id)
+        await db.increment_onboarding_runs(args.user_id)
         elapsed = asyncio.get_event_loop().time() - start_time
         logger.info(
             f"Graph execution started with ID {graph_exec.id} for graph {args.graph_id} "
@@ -246,8 +249,13 @@ def cleanup_expired_files():
 
 def cleanup_oauth_tokens():
     """Clean up expired OAuth tokens from the database."""
+
     # Wait for completion
-    run_async(cleanup_expired_oauth_tokens())
+    async def _cleanup():
+        db = get_database_manager_async_client()
+        return await db.cleanup_expired_oauth_tokens()
+
+    run_async(_cleanup())
 
 
 def execution_accuracy_alerts():

autogpt_platform/backend/backend/integrations/credentials_store.py

@@ -224,6 +224,14 @@ openweathermap_credentials = APIKeyCredentials(
     expires_at=None,
 )
 
+elevenlabs_credentials = APIKeyCredentials(
+    id="f4a8b6c2-3d1e-4f5a-9b8c-7d6e5f4a3b2c",
+    provider="elevenlabs",
+    api_key=SecretStr(settings.secrets.elevenlabs_api_key),
+    title="Use Credits for ElevenLabs",
+    expires_at=None,
+)
+
 DEFAULT_CREDENTIALS = [
     ollama_credentials,
     revid_credentials,
@@ -252,6 +260,7 @@ DEFAULT_CREDENTIALS = [
     v0_credentials,
     webshare_proxy_credentials,
     openweathermap_credentials,
+    elevenlabs_credentials,
 ]
 
 SYSTEM_CREDENTIAL_IDS = {cred.id for cred in DEFAULT_CREDENTIALS}
@@ -366,6 +375,8 @@ class IntegrationCredentialsStore:
             all_credentials.append(webshare_proxy_credentials)
         if settings.secrets.openweathermap_api_key:
             all_credentials.append(openweathermap_credentials)
+        if settings.secrets.elevenlabs_api_key:
+            all_credentials.append(elevenlabs_credentials)
         return all_credentials
 
     async def get_creds_by_id(

autogpt_platform/backend/backend/integrations/providers.py

@@ -18,6 +18,7 @@ class ProviderName(str, Enum):
     DISCORD = "discord"
     D_ID = "d_id"
     E2B = "e2b"
+    ELEVENLABS = "elevenlabs"
     FAL = "fal"
     GITHUB = "github"
     GOOGLE = "google"

autogpt_platform/backend/backend/integrations/webhooks/utils_test.py

@@ -0,0 +1,39 @@
+from urllib.parse import urlparse
+
+import fastapi
+from fastapi.routing import APIRoute
+
+from backend.api.features.integrations.router import router as integrations_router
+from backend.integrations.providers import ProviderName
+from backend.integrations.webhooks import utils as webhooks_utils
+
+
+def test_webhook_ingress_url_matches_route(monkeypatch) -> None:
+    app = fastapi.FastAPI()
+    app.include_router(integrations_router, prefix="/api/integrations")
+
+    provider = ProviderName.GITHUB
+    webhook_id = "webhook_123"
+    base_url = "https://example.com"
+
+    monkeypatch.setattr(webhooks_utils.app_config, "platform_base_url", base_url)
+
+    route = next(
+        route
+        for route in integrations_router.routes
+        if isinstance(route, APIRoute)
+        and route.path == "/{provider}/webhooks/{webhook_id}/ingress"
+        and "POST" in route.methods
+    )
+    expected_path = f"/api/integrations{route.path}".format(
+        provider=provider.value,
+        webhook_id=webhook_id,
+    )
+    actual_url = urlparse(webhooks_utils.webhook_ingress_url(provider, webhook_id))
+    expected_base = urlparse(base_url)
+
+    assert (actual_url.scheme, actual_url.netloc) == (
+        expected_base.scheme,
+        expected_base.netloc,
+    )
+    assert actual_url.path == expected_path

autogpt_platform/backend/backend/util/file.py

@@ -8,6 +8,8 @@ from pathlib import Path
 from typing import TYPE_CHECKING, Literal
 from urllib.parse import urlparse
 
+from pydantic import BaseModel
+
 from backend.util.cloud_storage import get_cloud_storage_handler
 from backend.util.request import Requests
 from backend.util.settings import Config
@@ -17,6 +19,35 @@ from backend.util.virus_scanner import scan_content_safe
 if TYPE_CHECKING:
     from backend.data.execution import ExecutionContext
 
+
+class WorkspaceUri(BaseModel):
+    """Parsed workspace:// URI."""
+
+    file_ref: str  # File ID or path (e.g. "abc123" or "/path/to/file.txt")
+    mime_type: str | None = None  # MIME type from fragment (e.g. "video/mp4")
+    is_path: bool = False  # True if file_ref is a path (starts with "/")
+
+
+def parse_workspace_uri(uri: str) -> WorkspaceUri:
+    """Parse a workspace:// URI into its components.
+
+    Examples:
+        "workspace://abc123"            → WorkspaceUri(file_ref="abc123", mime_type=None, is_path=False)
+        "workspace://abc123#video/mp4"  → WorkspaceUri(file_ref="abc123", mime_type="video/mp4", is_path=False)
+        "workspace:///path/to/file.txt" → WorkspaceUri(file_ref="/path/to/file.txt", mime_type=None, is_path=True)
+    """
+    raw = uri.removeprefix("workspace://")
+    mime_type: str | None = None
+    if "#" in raw:
+        raw, fragment = raw.split("#", 1)
+        mime_type = fragment or None
+    return WorkspaceUri(
+        file_ref=raw,
+        mime_type=mime_type,
+        is_path=raw.startswith("/"),
+    )
+
+
 # Return format options for store_media_file
 # - "for_local_processing": Returns local file path - use with ffmpeg, MoviePy, PIL, etc.
 # - "for_external_api": Returns data URI (base64) - use when sending content to external APIs
@@ -183,22 +214,20 @@ async def store_media_file(
                 "This file type is only available in CoPilot sessions."
             )
 
-        # Parse workspace reference
+        # Parse workspace reference (strips #mimeType fragment from file ID)
-        # workspace://abc123 - by file ID
+        ws = parse_workspace_uri(file)
-        # workspace:///path/to/file.txt - by virtual path
-        file_ref = file[12:]  # Remove "workspace://"
 
-        if file_ref.startswith("/"):
+        if ws.is_path:
-            # Path reference
+            # Path reference: workspace:///path/to/file.txt
-            workspace_content = await workspace_manager.read_file(file_ref)
+            workspace_content = await workspace_manager.read_file(ws.file_ref)
-            file_info = await workspace_manager.get_file_info_by_path(file_ref)
+            file_info = await workspace_manager.get_file_info_by_path(ws.file_ref)
             filename = sanitize_filename(
                 file_info.name if file_info else f"{uuid.uuid4()}.bin"
             )
         else:
-            # ID reference
+            # ID reference: workspace://abc123 or workspace://abc123#video/mp4
-            workspace_content = await workspace_manager.read_file_by_id(file_ref)
+            workspace_content = await workspace_manager.read_file_by_id(ws.file_ref)
-            file_info = await workspace_manager.get_file_info(file_ref)
+            file_info = await workspace_manager.get_file_info(ws.file_ref)
             filename = sanitize_filename(
                 file_info.name if file_info else f"{uuid.uuid4()}.bin"
             )
@@ -334,7 +363,21 @@ async def store_media_file(
 
         # Don't re-save if input was already from workspace
         if is_from_workspace:
-            # Return original workspace reference
+            # Return original workspace reference, ensuring MIME type fragment
+            ws = parse_workspace_uri(file)
+            if not ws.mime_type:
+                # Add MIME type fragment if missing (older refs without it)
+                try:
+                    if ws.is_path:
+                        info = await workspace_manager.get_file_info_by_path(
+                            ws.file_ref
+                        )
+                    else:
+                        info = await workspace_manager.get_file_info(ws.file_ref)
+                    if info:
+                        return MediaFileType(f"{file}#{info.mimeType}")
+                except Exception:
+                    pass
             return MediaFileType(file)
 
         # Save new content to workspace
@@ -346,7 +389,7 @@ async def store_media_file(
             filename=filename,
             overwrite=True,
         )
-        return MediaFileType(f"workspace://{file_record.id}")
+        return MediaFileType(f"workspace://{file_record.id}#{file_record.mimeType}")
 
     else:
         raise ValueError(f"Invalid return_format: {return_format}")

autogpt_platform/backend/backend/util/prompt.py

@@ -1,10 +1,19 @@
+from __future__ import annotations
+
+import logging
 from copy import deepcopy
-from typing import Any
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any
 
 from tiktoken import encoding_for_model
 
 from backend.util import json
 
+if TYPE_CHECKING:
+    from openai import AsyncOpenAI
+
+logger = logging.getLogger(__name__)
+
 # ---------------------------------------------------------------------------#
 #  CONSTANTS                                                                 #
 # ---------------------------------------------------------------------------#
@@ -100,9 +109,17 @@ def _is_objective_message(msg: dict) -> bool:
 def _truncate_tool_message_content(msg: dict, enc, max_tokens: int) -> None:
     """
     Carefully truncate tool message content while preserving tool structure.
-    Only truncates tool_result content, leaves tool_use intact.
+    Handles both Anthropic-style (list content) and OpenAI-style (string content) tool messages.
     """
     content = msg.get("content")
+
+    # OpenAI-style tool message: role="tool" with string content
+    if msg.get("role") == "tool" and isinstance(content, str):
+        if _tok_len(content, enc) > max_tokens:
+            msg["content"] = _truncate_middle_tokens(content, enc, max_tokens)
+        return
+
+    # Anthropic-style: list content with tool_result items
     if not isinstance(content, list):
         return
 
@@ -140,141 +157,6 @@ def _truncate_middle_tokens(text: str, enc, max_tok: int) -> str:
 # ---------------------------------------------------------------------------#
 
 
-def compress_prompt(
-    messages: list[dict],
-    target_tokens: int,
-    *,
-    model: str = "gpt-4o",
-    reserve: int = 2_048,
-    start_cap: int = 8_192,
-    floor_cap: int = 128,
-    lossy_ok: bool = True,
-) -> list[dict]:
-    """
-    Shrink *messages* so that::
-
-        token_count(prompt) + reserve  ≤  target_tokens
-
-    Strategy
-    --------
-    1. **Token-aware truncation** – progressively halve a per-message cap
-       (`start_cap`, `start_cap/2`, … `floor_cap`) and apply it to the
-       *content* of every message except the first and last.  Tool shells
-       are included: we keep the envelope but shorten huge payloads.
-    2. **Middle-out deletion** – if still over the limit, delete whole
-       messages working outward from the centre, **skipping** any message
-       that contains ``tool_calls`` or has ``role == "tool"``.
-    3. **Last-chance trim** – if still too big, truncate the *first* and
-       *last* message bodies down to `floor_cap` tokens.
-    4. If the prompt is *still* too large:
-         • raise ``ValueError``      when ``lossy_ok == False`` (default)
-         • return the partially-trimmed prompt when ``lossy_ok == True``
-
-    Parameters
-    ----------
-    messages        Complete chat history (will be deep-copied).
-    model           Model name; passed to tiktoken to pick the right
-                    tokenizer (gpt-4o → 'o200k_base', others fallback).
-    target_tokens   Hard ceiling for prompt size **excluding** the model's
-                    forthcoming answer.
-    reserve         How many tokens you want to leave available for that
-                    answer (`max_tokens` in your subsequent completion call).
-    start_cap       Initial per-message truncation ceiling (tokens).
-    floor_cap       Lowest cap we'll accept before moving to deletions.
-    lossy_ok        If *True* return best-effort prompt instead of raising
-                    after all trim passes have been exhausted.
-
-    Returns
-    -------
-    list[dict]  – A *new* messages list that abides by the rules above.
-    """
-    enc = encoding_for_model(model)  # best-match tokenizer
-    msgs = deepcopy(messages)  # never mutate caller
-
-    def total_tokens() -> int:
-        """Current size of *msgs* in tokens."""
-        return sum(_msg_tokens(m, enc) for m in msgs)
-
-    original_token_count = total_tokens()
-
-    if original_token_count + reserve <= target_tokens:
-        return msgs
-
-    # ---- STEP 0 : normalise content --------------------------------------
-    # Convert non-string payloads to strings so token counting is coherent.
-    for i, m in enumerate(msgs):
-        if not isinstance(m.get("content"), str) and m.get("content") is not None:
-            if _is_tool_message(m):
-                continue
-
-            # Keep first and last messages intact (unless they're tool messages)
-            if i == 0 or i == len(msgs) - 1:
-                continue
-
-            # Reasonable 20k-char ceiling prevents pathological blobs
-            content_str = json.dumps(m["content"], separators=(",", ":"))
-            if len(content_str) > 20_000:
-                content_str = _truncate_middle_tokens(content_str, enc, 20_000)
-            m["content"] = content_str
-
-    # ---- STEP 1 : token-aware truncation ---------------------------------
-    cap = start_cap
-    while total_tokens() + reserve > target_tokens and cap >= floor_cap:
-        for m in msgs[1:-1]:  # keep first & last intact
-            if _is_tool_message(m):
-                # For tool messages, only truncate tool result content, preserve structure
-                _truncate_tool_message_content(m, enc, cap)
-                continue
-
-            if _is_objective_message(m):
-                # Never truncate objective messages - they contain the core task
-                continue
-
-            content = m.get("content") or ""
-            if _tok_len(content, enc) > cap:
-                m["content"] = _truncate_middle_tokens(content, enc, cap)
-        cap //= 2  # tighten the screw
-
-    # ---- STEP 2 : middle-out deletion -----------------------------------
-    while total_tokens() + reserve > target_tokens and len(msgs) > 2:
-        # Identify all deletable messages (not first/last, not tool messages, not objective messages)
-        deletable_indices = []
-        for i in range(1, len(msgs) - 1):  # Skip first and last
-            if not _is_tool_message(msgs[i]) and not _is_objective_message(msgs[i]):
-                deletable_indices.append(i)
-
-        if not deletable_indices:
-            break  # nothing more we can drop
-
-        # Delete from center outward - find the index closest to center
-        centre = len(msgs) // 2
-        to_delete = min(deletable_indices, key=lambda i: abs(i - centre))
-        del msgs[to_delete]
-
-    # ---- STEP 3 : final safety-net trim on first & last ------------------
-    cap = start_cap
-    while total_tokens() + reserve > target_tokens and cap >= floor_cap:
-        for idx in (0, -1):  # first and last
-            if _is_tool_message(msgs[idx]):
-                # For tool messages at first/last position, truncate tool result content only
-                _truncate_tool_message_content(msgs[idx], enc, cap)
-                continue
-
-            text = msgs[idx].get("content") or ""
-            if _tok_len(text, enc) > cap:
-                msgs[idx]["content"] = _truncate_middle_tokens(text, enc, cap)
-        cap //= 2  # tighten the screw
-
-    # ---- STEP 4 : success or fail-gracefully -----------------------------
-    if total_tokens() + reserve > target_tokens and not lossy_ok:
-        raise ValueError(
-            "compress_prompt: prompt still exceeds budget "
-            f"({total_tokens() + reserve} > {target_tokens})."
-        )
-
-    return msgs
-
-
 def estimate_token_count(
     messages: list[dict],
     *,
@@ -293,7 +175,8 @@ def estimate_token_count(
     -------
     int  – Token count.
     """
-    enc = encoding_for_model(model)  # best-match tokenizer
+    token_model = _normalize_model_for_tokenizer(model)
+    enc = encoding_for_model(token_model)
     return sum(_msg_tokens(m, enc) for m in messages)
 
 
@@ -315,6 +198,543 @@ def estimate_token_count_str(
     -------
     int  – Token count.
     """
-    enc = encoding_for_model(model)  # best-match tokenizer
+    token_model = _normalize_model_for_tokenizer(model)
+    enc = encoding_for_model(token_model)
     text = json.dumps(text) if not isinstance(text, str) else text
     return _tok_len(text, enc)
+
+
+# ---------------------------------------------------------------------------#
+#  UNIFIED CONTEXT COMPRESSION                                               #
+# ---------------------------------------------------------------------------#
+
+# Default thresholds
+DEFAULT_TOKEN_THRESHOLD = 120_000
+DEFAULT_KEEP_RECENT = 15
+
+
+@dataclass
+class CompressResult:
+    """Result of context compression."""
+
+    messages: list[dict]
+    token_count: int
+    was_compacted: bool
+    error: str | None = None
+    original_token_count: int = 0
+    messages_summarized: int = 0
+    messages_dropped: int = 0
+
+
+def _normalize_model_for_tokenizer(model: str) -> str:
+    """Normalize model name for tiktoken tokenizer selection."""
+    if "/" in model:
+        model = model.split("/")[-1]
+    if "claude" in model.lower() or not any(
+        known in model.lower() for known in ["gpt", "o1", "chatgpt", "text-"]
+    ):
+        return "gpt-4o"
+    return model
+
+
+def _extract_tool_call_ids_from_message(msg: dict) -> set[str]:
+    """
+    Extract tool_call IDs from an assistant message.
+
+    Supports both formats:
+    - OpenAI: {"role": "assistant", "tool_calls": [{"id": "..."}]}
+    - Anthropic: {"role": "assistant", "content": [{"type": "tool_use", "id": "..."}]}
+
+    Returns:
+        Set of tool_call IDs found in the message.
+    """
+    ids: set[str] = set()
+    if msg.get("role") != "assistant":
+        return ids
+
+    # OpenAI format: tool_calls array
+    if msg.get("tool_calls"):
+        for tc in msg["tool_calls"]:
+            tc_id = tc.get("id")
+            if tc_id:
+                ids.add(tc_id)
+
+    # Anthropic format: content list with tool_use blocks
+    content = msg.get("content")
+    if isinstance(content, list):
+        for block in content:
+            if isinstance(block, dict) and block.get("type") == "tool_use":
+                tc_id = block.get("id")
+                if tc_id:
+                    ids.add(tc_id)
+
+    return ids
+
+
+def _extract_tool_response_ids_from_message(msg: dict) -> set[str]:
+    """
+    Extract tool_call IDs that this message is responding to.
+
+    Supports both formats:
+    - OpenAI: {"role": "tool", "tool_call_id": "..."}
+    - Anthropic: {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "..."}]}
+
+    Returns:
+        Set of tool_call IDs this message responds to.
+    """
+    ids: set[str] = set()
+
+    # OpenAI format: role=tool with tool_call_id
+    if msg.get("role") == "tool":
+        tc_id = msg.get("tool_call_id")
+        if tc_id:
+            ids.add(tc_id)
+
+    # Anthropic format: content list with tool_result blocks
+    content = msg.get("content")
+    if isinstance(content, list):
+        for block in content:
+            if isinstance(block, dict) and block.get("type") == "tool_result":
+                tc_id = block.get("tool_use_id")
+                if tc_id:
+                    ids.add(tc_id)
+
+    return ids
+
+
+def _is_tool_response_message(msg: dict) -> bool:
+    """Check if message is a tool response (OpenAI or Anthropic format)."""
+    # OpenAI format
+    if msg.get("role") == "tool":
+        return True
+    # Anthropic format
+    content = msg.get("content")
+    if isinstance(content, list):
+        for block in content:
+            if isinstance(block, dict) and block.get("type") == "tool_result":
+                return True
+    return False
+
+
+def _remove_orphan_tool_responses(
+    messages: list[dict], orphan_ids: set[str]
+) -> list[dict]:
+    """
+    Remove tool response messages/blocks that reference orphan tool_call IDs.
+
+    Supports both OpenAI and Anthropic formats.
+    For Anthropic messages with mixed valid/orphan tool_result blocks,
+    filters out only the orphan blocks instead of dropping the entire message.
+    """
+    result = []
+    for msg in messages:
+        # OpenAI format: role=tool - drop entire message if orphan
+        if msg.get("role") == "tool":
+            tc_id = msg.get("tool_call_id")
+            if tc_id and tc_id in orphan_ids:
+                continue
+            result.append(msg)
+            continue
+
+        # Anthropic format: content list may have mixed tool_result blocks
+        content = msg.get("content")
+        if isinstance(content, list):
+            has_tool_results = any(
+                isinstance(b, dict) and b.get("type") == "tool_result" for b in content
+            )
+            if has_tool_results:
+                # Filter out orphan tool_result blocks, keep valid ones
+                filtered_content = [
+                    block
+                    for block in content
+                    if not (
+                        isinstance(block, dict)
+                        and block.get("type") == "tool_result"
+                        and block.get("tool_use_id") in orphan_ids
+                    )
+                ]
+                # Only keep message if it has remaining content
+                if filtered_content:
+                    msg = msg.copy()
+                    msg["content"] = filtered_content
+                    result.append(msg)
+                continue
+
+        result.append(msg)
+    return result
+
+
+def _ensure_tool_pairs_intact(
+    recent_messages: list[dict],
+    all_messages: list[dict],
+    start_index: int,
+) -> list[dict]:
+    """
+    Ensure tool_call/tool_response pairs stay together after slicing.
+
+    When slicing messages for context compaction, a naive slice can separate
+    an assistant message containing tool_calls from its corresponding tool
+    response messages. This causes API validation errors (e.g., Anthropic's
+    "unexpected tool_use_id found in tool_result blocks").
+
+    This function checks for orphan tool responses in the slice and extends
+    backwards to include their corresponding assistant messages.
+
+    Supports both formats:
+    - OpenAI: tool_calls array + role="tool" responses
+    - Anthropic: tool_use blocks + tool_result blocks
+
+    Args:
+        recent_messages: The sliced messages to validate
+        all_messages: The complete message list (for looking up missing assistants)
+        start_index: The index in all_messages where recent_messages begins
+
+    Returns:
+        A potentially extended list of messages with tool pairs intact
+    """
+    if not recent_messages:
+        return recent_messages
+
+    # Collect all tool_call_ids from assistant messages in the slice
+    available_tool_call_ids: set[str] = set()
+    for msg in recent_messages:
+        available_tool_call_ids |= _extract_tool_call_ids_from_message(msg)
+
+    # Find orphan tool responses (responses whose tool_call_id is missing)
+    orphan_tool_call_ids: set[str] = set()
+    for msg in recent_messages:
+        response_ids = _extract_tool_response_ids_from_message(msg)
+        for tc_id in response_ids:
+            if tc_id not in available_tool_call_ids:
+                orphan_tool_call_ids.add(tc_id)
+
+    if not orphan_tool_call_ids:
+        # No orphans, slice is valid
+        return recent_messages
+
+    # Find the assistant messages that contain the orphan tool_call_ids
+    # Search backwards from start_index in all_messages
+    messages_to_prepend: list[dict] = []
+    for i in range(start_index - 1, -1, -1):
+        msg = all_messages[i]
+        msg_tool_ids = _extract_tool_call_ids_from_message(msg)
+        if msg_tool_ids & orphan_tool_call_ids:
+            # This assistant message has tool_calls we need
+            # Also collect its contiguous tool responses that follow it
+            assistant_and_responses: list[dict] = [msg]
+
+            # Scan forward from this assistant to collect tool responses
+            for j in range(i + 1, start_index):
+                following_msg = all_messages[j]
+                following_response_ids = _extract_tool_response_ids_from_message(
+                    following_msg
+                )
+                if following_response_ids and following_response_ids & msg_tool_ids:
+                    assistant_and_responses.append(following_msg)
+                elif not _is_tool_response_message(following_msg):
+                    # Stop at first non-tool-response message
+                    break
+
+            # Prepend the assistant and its tool responses (maintain order)
+            messages_to_prepend = assistant_and_responses + messages_to_prepend
+            # Mark these as found
+            orphan_tool_call_ids -= msg_tool_ids
+            # Also add this assistant's tool_call_ids to available set
+            available_tool_call_ids |= msg_tool_ids
+
+        if not orphan_tool_call_ids:
+            # Found all missing assistants
+            break
+
+    if orphan_tool_call_ids:
+        # Some tool_call_ids couldn't be resolved - remove those tool responses
+        # This shouldn't happen in normal operation but handles edge cases
+        logger.warning(
+            f"Could not find assistant messages for tool_call_ids: {orphan_tool_call_ids}. "
+            "Removing orphan tool responses."
+        )
+        recent_messages = _remove_orphan_tool_responses(
+            recent_messages, orphan_tool_call_ids
+        )
+
+    if messages_to_prepend:
+        logger.info(
+            f"Extended recent messages by {len(messages_to_prepend)} to preserve "
+            f"tool_call/tool_response pairs"
+        )
+        return messages_to_prepend + recent_messages
+
+    return recent_messages
+
+
+async def _summarize_messages_llm(
+    messages: list[dict],
+    client: AsyncOpenAI,
+    model: str,
+    timeout: float = 30.0,
+) -> str:
+    """Summarize messages using an LLM."""
+    conversation = []
+    for msg in messages:
+        role = msg.get("role", "")
+        content = msg.get("content", "")
+        if content and role in ("user", "assistant", "tool"):
+            conversation.append(f"{role.upper()}: {content}")
+
+    conversation_text = "\n\n".join(conversation)
+
+    if not conversation_text:
+        return "No conversation history available."
+
+    # Limit to ~100k chars for safety
+    MAX_CHARS = 100_000
+    if len(conversation_text) > MAX_CHARS:
+        conversation_text = conversation_text[:MAX_CHARS] + "\n\n[truncated]"
+
+    response = await client.with_options(timeout=timeout).chat.completions.create(
+        model=model,
+        messages=[
+            {
+                "role": "system",
+                "content": (
+                    "Create a detailed summary of the conversation so far. "
+                    "This summary will be used as context when continuing the conversation.\n\n"
+                    "Before writing the summary, analyze each message chronologically to identify:\n"
+                    "- User requests and their explicit goals\n"
+                    "- Your approach and key decisions made\n"
+                    "- Technical specifics (file names, tool outputs, function signatures)\n"
+                    "- Errors encountered and resolutions applied\n\n"
+                    "You MUST include ALL of the following sections:\n\n"
+                    "## 1. Primary Request and Intent\n"
+                    "The user's explicit goals and what they are trying to accomplish.\n\n"
+                    "## 2. Key Technical Concepts\n"
+                    "Technologies, frameworks, tools, and patterns being used or discussed.\n\n"
+                    "## 3. Files and Resources Involved\n"
+                    "Specific files examined or modified, with relevant snippets and identifiers.\n\n"
+                    "## 4. Errors and Fixes\n"
+                    "Problems encountered, error messages, and their resolutions. "
+                    "Include any user feedback on fixes.\n\n"
+                    "## 5. Problem Solving\n"
+                    "Issues that have been resolved and how they were addressed.\n\n"
+                    "## 6. All User Messages\n"
+                    "A complete list of all user inputs (excluding tool outputs) to preserve their exact requests.\n\n"
+                    "## 7. Pending Tasks\n"
+                    "Work items the user explicitly requested that have not yet been completed.\n\n"
+                    "## 8. Current Work\n"
+                    "Precise description of what was being worked on most recently, including relevant context.\n\n"
+                    "## 9. Next Steps\n"
+                    "What should happen next, aligned with the user's most recent requests. "
+                    "Include verbatim quotes of recent instructions if relevant."
+                ),
+            },
+            {"role": "user", "content": f"Summarize:\n\n{conversation_text}"},
+        ],
+        max_tokens=1500,
+        temperature=0.3,
+    )
+
+    return response.choices[0].message.content or "No summary available."
+
+
+async def compress_context(
+    messages: list[dict],
+    target_tokens: int = DEFAULT_TOKEN_THRESHOLD,
+    *,
+    model: str = "gpt-4o",
+    client: AsyncOpenAI | None = None,
+    keep_recent: int = DEFAULT_KEEP_RECENT,
+    reserve: int = 2_048,
+    start_cap: int = 8_192,
+    floor_cap: int = 128,
+) -> CompressResult:
+    """
+    Unified context compression that combines summarization and truncation strategies.
+
+    Strategy (in order):
+    1. **LLM summarization** – If client provided, summarize old messages into a
+       single context message while keeping recent messages intact. This is the
+       primary strategy for chat service.
+    2. **Content truncation** – Progressively halve a per-message cap and truncate
+       bloated message content (tool outputs, large pastes). Preserves all messages
+       but shortens their content. Primary strategy when client=None (LLM blocks).
+    3. **Middle-out deletion** – Delete whole messages one at a time from the center
+       outward, skipping tool messages and objective messages.
+    4. **First/last trim** – Truncate first and last message content as last resort.
+
+    Parameters
+    ----------
+    messages        Complete chat history (will be deep-copied).
+    target_tokens   Hard ceiling for prompt size.
+    model           Model name for tokenization and summarization.
+    client          AsyncOpenAI client. If provided, enables LLM summarization
+                    as the first strategy. If None, skips to truncation strategies.
+    keep_recent     Number of recent messages to preserve during summarization.
+    reserve         Tokens to reserve for model response.
+    start_cap       Initial per-message truncation ceiling (tokens).
+    floor_cap       Lowest cap before moving to deletions.
+
+    Returns
+    -------
+    CompressResult with compressed messages and metadata.
+    """
+    # Guard clause for empty messages
+    if not messages:
+        return CompressResult(
+            messages=[],
+            token_count=0,
+            was_compacted=False,
+            original_token_count=0,
+        )
+
+    token_model = _normalize_model_for_tokenizer(model)
+    enc = encoding_for_model(token_model)
+    msgs = deepcopy(messages)
+
+    def total_tokens() -> int:
+        return sum(_msg_tokens(m, enc) for m in msgs)
+
+    original_count = total_tokens()
+
+    # Already under limit
+    if original_count + reserve <= target_tokens:
+        return CompressResult(
+            messages=msgs,
+            token_count=original_count,
+            was_compacted=False,
+            original_token_count=original_count,
+        )
+
+    messages_summarized = 0
+    messages_dropped = 0
+
+    # ---- STEP 1: LLM summarization (if client provided) -------------------
+    # This is the primary compression strategy for chat service.
+    # Summarize old messages while keeping recent ones intact.
+    if client is not None:
+        has_system = len(msgs) > 0 and msgs[0].get("role") == "system"
+        system_msg = msgs[0] if has_system else None
+
+        # Calculate old vs recent messages
+        if has_system:
+            if len(msgs) > keep_recent + 1:
+                old_msgs = msgs[1:-keep_recent]
+                recent_msgs = msgs[-keep_recent:]
+            else:
+                old_msgs = []
+                recent_msgs = msgs[1:] if len(msgs) > 1 else []
+        else:
+            if len(msgs) > keep_recent:
+                old_msgs = msgs[:-keep_recent]
+                recent_msgs = msgs[-keep_recent:]
+            else:
+                old_msgs = []
+                recent_msgs = msgs
+
+        # Ensure tool pairs stay intact
+        slice_start = max(0, len(msgs) - keep_recent)
+        recent_msgs = _ensure_tool_pairs_intact(recent_msgs, msgs, slice_start)
+
+        if old_msgs:
+            try:
+                summary_text = await _summarize_messages_llm(old_msgs, client, model)
+                summary_msg = {
+                    "role": "assistant",
+                    "content": f"[Previous conversation summary — for context only]: {summary_text}",
+                }
+                messages_summarized = len(old_msgs)
+
+                if has_system:
+                    msgs = [system_msg, summary_msg] + recent_msgs
+                else:
+                    msgs = [summary_msg] + recent_msgs
+
+                logger.info(
+                    f"Context summarized: {original_count} -> {total_tokens()} tokens, "
+                    f"summarized {messages_summarized} messages"
+                )
+            except Exception as e:
+                logger.warning(f"Summarization failed, continuing with truncation: {e}")
+                # Fall through to content truncation
+
+    # ---- STEP 2: Normalize content ----------------------------------------
+    # Convert non-string payloads to strings so token counting is coherent.
+    # Always run this before truncation to ensure consistent token counting.
+    for i, m in enumerate(msgs):
+        if not isinstance(m.get("content"), str) and m.get("content") is not None:
+            if _is_tool_message(m):
+                continue
+            if i == 0 or i == len(msgs) - 1:
+                continue
+            content_str = json.dumps(m["content"], separators=(",", ":"))
+            if len(content_str) > 20_000:
+                content_str = _truncate_middle_tokens(content_str, enc, 20_000)
+            m["content"] = content_str
+
+    # ---- STEP 3: Token-aware content truncation ---------------------------
+    # Progressively halve per-message cap and truncate bloated content.
+    # This preserves all messages but shortens their content.
+    cap = start_cap
+    while total_tokens() + reserve > target_tokens and cap >= floor_cap:
+        for m in msgs[1:-1]:
+            if _is_tool_message(m):
+                _truncate_tool_message_content(m, enc, cap)
+                continue
+            if _is_objective_message(m):
+                continue
+            content = m.get("content") or ""
+            if _tok_len(content, enc) > cap:
+                m["content"] = _truncate_middle_tokens(content, enc, cap)
+        cap //= 2
+
+    # ---- STEP 4: Middle-out deletion --------------------------------------
+    # Delete messages one at a time from the center outward.
+    # This is more granular than dropping all old messages at once.
+    while total_tokens() + reserve > target_tokens and len(msgs) > 2:
+        deletable: list[int] = []
+        for i in range(1, len(msgs) - 1):
+            msg = msgs[i]
+            if (
+                msg is not None
+                and not _is_tool_message(msg)
+                and not _is_objective_message(msg)
+            ):
+                deletable.append(i)
+        if not deletable:
+            break
+        centre = len(msgs) // 2
+        to_delete = min(deletable, key=lambda i: abs(i - centre))
+        del msgs[to_delete]
+        messages_dropped += 1
+
+    # ---- STEP 5: Final trim on first/last ---------------------------------
+    cap = start_cap
+    while total_tokens() + reserve > target_tokens and cap >= floor_cap:
+        for idx in (0, -1):
+            msg = msgs[idx]
+            if msg is None:
+                continue
+            if _is_tool_message(msg):
+                _truncate_tool_message_content(msg, enc, cap)
+                continue
+            text = msg.get("content") or ""
+            if _tok_len(text, enc) > cap:
+                msg["content"] = _truncate_middle_tokens(text, enc, cap)
+        cap //= 2
+
+    # Filter out any None values that may have been introduced
+    final_msgs: list[dict] = [m for m in msgs if m is not None]
+    final_count = sum(_msg_tokens(m, enc) for m in final_msgs)
+    error = None
+    if final_count + reserve > target_tokens:
+        error = f"Could not compress below target ({final_count + reserve} > {target_tokens})"
+        logger.warning(error)
+
+    return CompressResult(
+        messages=final_msgs,
+        token_count=final_count,
+        was_compacted=True,
+        error=error,
+        original_token_count=original_count,
+        messages_summarized=messages_summarized,
+        messages_dropped=messages_dropped,
+    )

autogpt_platform/backend/backend/util/prompt_test.py

@@ -1,10 +1,21 @@
 """Tests for prompt utility functions, especially tool call token counting."""
 
+from unittest.mock import AsyncMock, MagicMock
+
 import pytest
 from tiktoken import encoding_for_model
 
 from backend.util import json
-from backend.util.prompt import _msg_tokens, estimate_token_count
+from backend.util.prompt import (
+    CompressResult,
+    _ensure_tool_pairs_intact,
+    _msg_tokens,
+    _normalize_model_for_tokenizer,
+    _truncate_middle_tokens,
+    _truncate_tool_message_content,
+    compress_context,
+    estimate_token_count,
+)
 
 
 class TestMsgTokens:
@@ -276,3 +287,690 @@ class TestEstimateTokenCount:
 
         assert total_tokens == expected_total
         assert total_tokens > 20  # Should be substantial
+
+
+class TestNormalizeModelForTokenizer:
+    """Test model name normalization for tiktoken."""
+
+    def test_openai_models_unchanged(self):
+        """Test that OpenAI models are returned as-is."""
+        assert _normalize_model_for_tokenizer("gpt-4o") == "gpt-4o"
+        assert _normalize_model_for_tokenizer("gpt-4") == "gpt-4"
+        assert _normalize_model_for_tokenizer("gpt-3.5-turbo") == "gpt-3.5-turbo"
+
+    def test_claude_models_normalized(self):
+        """Test that Claude models are normalized to gpt-4o."""
+        assert _normalize_model_for_tokenizer("claude-3-opus") == "gpt-4o"
+        assert _normalize_model_for_tokenizer("claude-3-sonnet") == "gpt-4o"
+        assert _normalize_model_for_tokenizer("anthropic/claude-3-haiku") == "gpt-4o"
+
+    def test_openrouter_paths_extracted(self):
+        """Test that OpenRouter model paths are handled."""
+        assert _normalize_model_for_tokenizer("openai/gpt-4o") == "gpt-4o"
+        assert _normalize_model_for_tokenizer("anthropic/claude-3-opus") == "gpt-4o"
+
+    def test_unknown_models_default_to_gpt4o(self):
+        """Test that unknown models default to gpt-4o."""
+        assert _normalize_model_for_tokenizer("some-random-model") == "gpt-4o"
+        assert _normalize_model_for_tokenizer("llama-3-70b") == "gpt-4o"
+
+
+class TestTruncateToolMessageContent:
+    """Test tool message content truncation."""
+
+    @pytest.fixture
+    def enc(self):
+        return encoding_for_model("gpt-4o")
+
+    def test_truncate_openai_tool_message(self, enc):
+        """Test truncation of OpenAI-style tool message with string content."""
+        long_content = "x" * 10000
+        msg = {"role": "tool", "tool_call_id": "call_123", "content": long_content}
+
+        _truncate_tool_message_content(msg, enc, max_tokens=100)
+
+        # Content should be truncated
+        assert len(msg["content"]) < len(long_content)
+        assert "…" in msg["content"]  # Has ellipsis marker
+
+    def test_truncate_anthropic_tool_result(self, enc):
+        """Test truncation of Anthropic-style tool_result."""
+        long_content = "y" * 10000
+        msg = {
+            "role": "user",
+            "content": [
+                {
+                    "type": "tool_result",
+                    "tool_use_id": "toolu_123",
+                    "content": long_content,
+                }
+            ],
+        }
+
+        _truncate_tool_message_content(msg, enc, max_tokens=100)
+
+        # Content should be truncated
+        result_content = msg["content"][0]["content"]
+        assert len(result_content) < len(long_content)
+        assert "…" in result_content
+
+    def test_preserve_tool_use_blocks(self, enc):
+        """Test that tool_use blocks are not truncated."""
+        msg = {
+            "role": "assistant",
+            "content": [
+                {
+                    "type": "tool_use",
+                    "id": "toolu_123",
+                    "name": "some_function",
+                    "input": {"key": "value" * 1000},  # Large input
+                }
+            ],
+        }
+
+        original = json.dumps(msg["content"][0]["input"])
+        _truncate_tool_message_content(msg, enc, max_tokens=10)
+
+        # tool_use should be unchanged
+        assert json.dumps(msg["content"][0]["input"]) == original
+
+    def test_no_truncation_when_under_limit(self, enc):
+        """Test that short content is not modified."""
+        msg = {"role": "tool", "tool_call_id": "call_123", "content": "Short content"}
+
+        original = msg["content"]
+        _truncate_tool_message_content(msg, enc, max_tokens=1000)
+
+        assert msg["content"] == original
+
+
+class TestTruncateMiddleTokens:
+    """Test middle truncation of text."""
+
+    @pytest.fixture
+    def enc(self):
+        return encoding_for_model("gpt-4o")
+
+    def test_truncates_long_text(self, enc):
+        """Test that long text is truncated with ellipsis in middle."""
+        long_text = "word " * 1000
+        result = _truncate_middle_tokens(long_text, enc, max_tok=50)
+
+        assert len(enc.encode(result)) <= 52  # Allow some slack for ellipsis
+        assert "…" in result
+        assert result.startswith("word")  # Head preserved
+        assert result.endswith("word ")  # Tail preserved
+
+    def test_preserves_short_text(self, enc):
+        """Test that short text is not modified."""
+        short_text = "Hello world"
+        result = _truncate_middle_tokens(short_text, enc, max_tok=100)
+
+        assert result == short_text
+
+
+class TestEnsureToolPairsIntact:
+    """Test tool call/response pair preservation for both OpenAI and Anthropic formats."""
+
+    # ---- OpenAI Format Tests ----
+
+    def test_openai_adds_missing_tool_call(self):
+        """Test that orphaned OpenAI tool_response gets its tool_call prepended."""
+        all_msgs = [
+            {"role": "system", "content": "You are helpful."},
+            {
+                "role": "assistant",
+                "tool_calls": [
+                    {"id": "call_1", "type": "function", "function": {"name": "f1"}}
+                ],
+            },
+            {"role": "tool", "tool_call_id": "call_1", "content": "result"},
+            {"role": "user", "content": "Thanks!"},
+        ]
+        # Recent messages start at index 2 (the tool response)
+        recent = [all_msgs[2], all_msgs[3]]
+        start_index = 2
+
+        result = _ensure_tool_pairs_intact(recent, all_msgs, start_index)
+
+        # Should prepend the tool_call message
+        assert len(result) == 3
+        assert result[0]["role"] == "assistant"
+        assert "tool_calls" in result[0]
+
+    def test_openai_keeps_complete_pairs(self):
+        """Test that complete OpenAI pairs are unchanged."""
+        all_msgs = [
+            {"role": "system", "content": "System"},
+            {
+                "role": "assistant",
+                "tool_calls": [
+                    {"id": "call_1", "type": "function", "function": {"name": "f1"}}
+                ],
+            },
+            {"role": "tool", "tool_call_id": "call_1", "content": "result"},
+        ]
+        recent = all_msgs[1:]  # Include both tool_call and response
+        start_index = 1
+
+        result = _ensure_tool_pairs_intact(recent, all_msgs, start_index)
+
+        assert len(result) == 2  # No messages added
+
+    def test_openai_multiple_tool_calls(self):
+        """Test multiple OpenAI tool calls in one assistant message."""
+        all_msgs = [
+            {"role": "system", "content": "System"},
+            {
+                "role": "assistant",
+                "tool_calls": [
+                    {"id": "call_1", "type": "function", "function": {"name": "f1"}},
+                    {"id": "call_2", "type": "function", "function": {"name": "f2"}},
+                ],
+            },
+            {"role": "tool", "tool_call_id": "call_1", "content": "result1"},
+            {"role": "tool", "tool_call_id": "call_2", "content": "result2"},
+            {"role": "user", "content": "Thanks!"},
+        ]
+        # Recent messages start at index 2 (first tool response)
+        recent = [all_msgs[2], all_msgs[3], all_msgs[4]]
+        start_index = 2
+
+        result = _ensure_tool_pairs_intact(recent, all_msgs, start_index)
+
+        # Should prepend the assistant message with both tool_calls
+        assert len(result) == 4
+        assert result[0]["role"] == "assistant"
+        assert len(result[0]["tool_calls"]) == 2
+
+    # ---- Anthropic Format Tests ----
+
+    def test_anthropic_adds_missing_tool_use(self):
+        """Test that orphaned Anthropic tool_result gets its tool_use prepended."""
+        all_msgs = [
+            {"role": "system", "content": "You are helpful."},
+            {
+                "role": "assistant",
+                "content": [
+                    {
+                        "type": "tool_use",
+                        "id": "toolu_123",
+                        "name": "get_weather",
+                        "input": {"location": "SF"},
+                    }
+                ],
+            },
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_123",
+                        "content": "22°C and sunny",
+                    }
+                ],
+            },
+            {"role": "user", "content": "Thanks!"},
+        ]
+        # Recent messages start at index 2 (the tool_result)
+        recent = [all_msgs[2], all_msgs[3]]
+        start_index = 2
+
+        result = _ensure_tool_pairs_intact(recent, all_msgs, start_index)
+
+        # Should prepend the tool_use message
+        assert len(result) == 3
+        assert result[0]["role"] == "assistant"
+        assert result[0]["content"][0]["type"] == "tool_use"
+
+    def test_anthropic_keeps_complete_pairs(self):
+        """Test that complete Anthropic pairs are unchanged."""
+        all_msgs = [
+            {"role": "system", "content": "System"},
+            {
+                "role": "assistant",
+                "content": [
+                    {
+                        "type": "tool_use",
+                        "id": "toolu_456",
+                        "name": "calculator",
+                        "input": {"expr": "2+2"},
+                    }
+                ],
+            },
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_456",
+                        "content": "4",
+                    }
+                ],
+            },
+        ]
+        recent = all_msgs[1:]  # Include both tool_use and result
+        start_index = 1
+
+        result = _ensure_tool_pairs_intact(recent, all_msgs, start_index)
+
+        assert len(result) == 2  # No messages added
+
+    def test_anthropic_multiple_tool_uses(self):
+        """Test multiple Anthropic tool_use blocks in one message."""
+        all_msgs = [
+            {"role": "system", "content": "System"},
+            {
+                "role": "assistant",
+                "content": [
+                    {"type": "text", "text": "Let me check both..."},
+                    {
+                        "type": "tool_use",
+                        "id": "toolu_1",
+                        "name": "get_weather",
+                        "input": {"city": "NYC"},
+                    },
+                    {
+                        "type": "tool_use",
+                        "id": "toolu_2",
+                        "name": "get_weather",
+                        "input": {"city": "LA"},
+                    },
+                ],
+            },
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_1",
+                        "content": "Cold",
+                    },
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_2",
+                        "content": "Warm",
+                    },
+                ],
+            },
+            {"role": "user", "content": "Thanks!"},
+        ]
+        # Recent messages start at index 2 (tool_result)
+        recent = [all_msgs[2], all_msgs[3]]
+        start_index = 2
+
+        result = _ensure_tool_pairs_intact(recent, all_msgs, start_index)
+
+        # Should prepend the assistant message with both tool_uses
+        assert len(result) == 3
+        assert result[0]["role"] == "assistant"
+        tool_use_count = sum(
+            1 for b in result[0]["content"] if b.get("type") == "tool_use"
+        )
+        assert tool_use_count == 2
+
+    # ---- Mixed/Edge Case Tests ----
+
+    def test_anthropic_with_type_message_field(self):
+        """Test Anthropic format with 'type': 'message' field (smart_decision_maker style)."""
+        all_msgs = [
+            {"role": "system", "content": "You are helpful."},
+            {
+                "role": "assistant",
+                "content": [
+                    {
+                        "type": "tool_use",
+                        "id": "toolu_abc",
+                        "name": "search",
+                        "input": {"q": "test"},
+                    }
+                ],
+            },
+            {
+                "role": "user",
+                "type": "message",  # Extra field from smart_decision_maker
+                "content": [
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_abc",
+                        "content": "Found results",
+                    }
+                ],
+            },
+            {"role": "user", "content": "Thanks!"},
+        ]
+        # Recent messages start at index 2 (the tool_result with 'type': 'message')
+        recent = [all_msgs[2], all_msgs[3]]
+        start_index = 2
+
+        result = _ensure_tool_pairs_intact(recent, all_msgs, start_index)
+
+        # Should prepend the tool_use message
+        assert len(result) == 3
+        assert result[0]["role"] == "assistant"
+        assert result[0]["content"][0]["type"] == "tool_use"
+
+    def test_handles_no_tool_messages(self):
+        """Test messages without tool calls."""
+        all_msgs = [
+            {"role": "user", "content": "Hello"},
+            {"role": "assistant", "content": "Hi there!"},
+        ]
+        recent = all_msgs
+        start_index = 0
+
+        result = _ensure_tool_pairs_intact(recent, all_msgs, start_index)
+
+        assert result == all_msgs
+
+    def test_handles_empty_messages(self):
+        """Test empty message list."""
+        result = _ensure_tool_pairs_intact([], [], 0)
+        assert result == []
+
+    def test_mixed_text_and_tool_content(self):
+        """Test Anthropic message with mixed text and tool_use content."""
+        all_msgs = [
+            {
+                "role": "assistant",
+                "content": [
+                    {"type": "text", "text": "I'll help you with that."},
+                    {
+                        "type": "tool_use",
+                        "id": "toolu_mixed",
+                        "name": "search",
+                        "input": {"q": "test"},
+                    },
+                ],
+            },
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_mixed",
+                        "content": "Found results",
+                    }
+                ],
+            },
+            {"role": "assistant", "content": "Here are the results..."},
+        ]
+        # Start from tool_result
+        recent = [all_msgs[1], all_msgs[2]]
+        start_index = 1
+
+        result = _ensure_tool_pairs_intact(recent, all_msgs, start_index)
+
+        # Should prepend the assistant message with tool_use
+        assert len(result) == 3
+        assert result[0]["content"][0]["type"] == "text"
+        assert result[0]["content"][1]["type"] == "tool_use"
+
+
+class TestCompressContext:
+    """Test the async compress_context function."""
+
+    @pytest.mark.asyncio
+    async def test_no_compression_needed(self):
+        """Test messages under limit return without compression."""
+        messages = [
+            {"role": "system", "content": "You are helpful."},
+            {"role": "user", "content": "Hello!"},
+        ]
+
+        result = await compress_context(messages, target_tokens=100000)
+
+        assert isinstance(result, CompressResult)
+        assert result.was_compacted is False
+        assert len(result.messages) == 2
+        assert result.error is None
+
+    @pytest.mark.asyncio
+    async def test_truncation_without_client(self):
+        """Test that truncation works without LLM client."""
+        long_content = "x" * 50000
+        messages = [
+            {"role": "system", "content": "System"},
+            {"role": "user", "content": long_content},
+            {"role": "assistant", "content": "Response"},
+        ]
+
+        result = await compress_context(
+            messages, target_tokens=1000, client=None, reserve=100
+        )
+
+        assert result.was_compacted is True
+        # Should have truncated without summarization
+        assert result.messages_summarized == 0
+
+    @pytest.mark.asyncio
+    async def test_with_mocked_llm_client(self):
+        """Test summarization with mocked LLM client."""
+        # Create many messages to trigger summarization
+        messages = [{"role": "system", "content": "System prompt"}]
+        for i in range(30):
+            messages.append({"role": "user", "content": f"User message {i} " * 100})
+            messages.append(
+                {"role": "assistant", "content": f"Assistant response {i} " * 100}
+            )
+
+        # Mock the AsyncOpenAI client
+        mock_client = AsyncMock()
+        mock_response = MagicMock()
+        mock_response.choices = [MagicMock()]
+        mock_response.choices[0].message.content = "Summary of conversation"
+        mock_client.with_options.return_value.chat.completions.create = AsyncMock(
+            return_value=mock_response
+        )
+
+        result = await compress_context(
+            messages,
+            target_tokens=5000,
+            client=mock_client,
+            keep_recent=5,
+            reserve=500,
+        )
+
+        assert result.was_compacted is True
+        # Should have attempted summarization
+        assert mock_client.with_options.called or result.messages_summarized > 0
+
+    @pytest.mark.asyncio
+    async def test_preserves_tool_pairs(self):
+        """Test that tool call/response pairs stay together."""
+        messages = [
+            {"role": "system", "content": "System"},
+            {"role": "user", "content": "Do something"},
+            {
+                "role": "assistant",
+                "tool_calls": [
+                    {"id": "call_1", "type": "function", "function": {"name": "func"}}
+                ],
+            },
+            {"role": "tool", "tool_call_id": "call_1", "content": "Result " * 1000},
+            {"role": "assistant", "content": "Done!"},
+        ]
+
+        result = await compress_context(
+            messages, target_tokens=500, client=None, reserve=50
+        )
+
+        # Check that if tool response exists, its call exists too
+        tool_call_ids = set()
+        tool_response_ids = set()
+        for msg in result.messages:
+            if "tool_calls" in msg:
+                for tc in msg["tool_calls"]:
+                    tool_call_ids.add(tc["id"])
+            if msg.get("role") == "tool":
+                tool_response_ids.add(msg.get("tool_call_id"))
+
+        # All tool responses should have their calls
+        assert tool_response_ids <= tool_call_ids
+
+    @pytest.mark.asyncio
+    async def test_returns_error_when_cannot_compress(self):
+        """Test that error is returned when compression fails."""
+        # Single huge message that can't be compressed enough
+        messages = [
+            {"role": "user", "content": "x" * 100000},
+        ]
+
+        result = await compress_context(
+            messages, target_tokens=100, client=None, reserve=50
+        )
+
+        # Should have an error since we can't get below 100 tokens
+        assert result.error is not None
+        assert result.was_compacted is True
+
+    @pytest.mark.asyncio
+    async def test_empty_messages(self):
+        """Test that empty messages list returns early without error."""
+        result = await compress_context([], target_tokens=1000)
+
+        assert result.messages == []
+        assert result.token_count == 0
+        assert result.was_compacted is False
+        assert result.error is None
+
+
+class TestRemoveOrphanToolResponses:
+    """Test _remove_orphan_tool_responses helper function."""
+
+    def test_removes_openai_orphan(self):
+        """Test removal of orphan OpenAI tool response."""
+        from backend.util.prompt import _remove_orphan_tool_responses
+
+        messages = [
+            {"role": "tool", "tool_call_id": "call_orphan", "content": "result"},
+            {"role": "user", "content": "Hello"},
+        ]
+        orphan_ids = {"call_orphan"}
+
+        result = _remove_orphan_tool_responses(messages, orphan_ids)
+
+        assert len(result) == 1
+        assert result[0]["role"] == "user"
+
+    def test_keeps_valid_openai_tool(self):
+        """Test that valid OpenAI tool responses are kept."""
+        from backend.util.prompt import _remove_orphan_tool_responses
+
+        messages = [
+            {"role": "tool", "tool_call_id": "call_valid", "content": "result"},
+        ]
+        orphan_ids = {"call_other"}
+
+        result = _remove_orphan_tool_responses(messages, orphan_ids)
+
+        assert len(result) == 1
+        assert result[0]["tool_call_id"] == "call_valid"
+
+    def test_filters_anthropic_mixed_blocks(self):
+        """Test filtering individual orphan blocks from Anthropic message with mixed valid/orphan."""
+        from backend.util.prompt import _remove_orphan_tool_responses
+
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_valid",
+                        "content": "valid result",
+                    },
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_orphan",
+                        "content": "orphan result",
+                    },
+                ],
+            },
+        ]
+        orphan_ids = {"toolu_orphan"}
+
+        result = _remove_orphan_tool_responses(messages, orphan_ids)
+
+        assert len(result) == 1
+        # Should only have the valid tool_result, orphan filtered out
+        assert len(result[0]["content"]) == 1
+        assert result[0]["content"][0]["tool_use_id"] == "toolu_valid"
+
+    def test_removes_anthropic_all_orphan(self):
+        """Test removal of Anthropic message when all tool_results are orphans."""
+        from backend.util.prompt import _remove_orphan_tool_responses
+
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_orphan1",
+                        "content": "result1",
+                    },
+                    {
+                        "type": "tool_result",
+                        "tool_use_id": "toolu_orphan2",
+                        "content": "result2",
+                    },
+                ],
+            },
+        ]
+        orphan_ids = {"toolu_orphan1", "toolu_orphan2"}
+
+        result = _remove_orphan_tool_responses(messages, orphan_ids)
+
+        # Message should be completely removed since no content left
+        assert len(result) == 0
+
+    def test_preserves_non_tool_messages(self):
+        """Test that non-tool messages are preserved."""
+        from backend.util.prompt import _remove_orphan_tool_responses
+
+        messages = [
+            {"role": "user", "content": "Hello"},
+            {"role": "assistant", "content": "Hi there!"},
+        ]
+        orphan_ids = {"some_id"}
+
+        result = _remove_orphan_tool_responses(messages, orphan_ids)
+
+        assert result == messages
+
+
+class TestCompressResultDataclass:
+    """Test CompressResult dataclass."""
+
+    def test_default_values(self):
+        """Test default values are set correctly."""
+        result = CompressResult(
+            messages=[{"role": "user", "content": "test"}],
+            token_count=10,
+            was_compacted=False,
+        )
+
+        assert result.error is None
+        assert result.original_token_count == 0  # Defaults to 0, not None
+        assert result.messages_summarized == 0
+        assert result.messages_dropped == 0
+
+    def test_all_fields(self):
+        """Test all fields can be set."""
+        result = CompressResult(
+            messages=[{"role": "user", "content": "test"}],
+            token_count=100,
+            was_compacted=True,
+            error="Some error",
+            original_token_count=500,
+            messages_summarized=10,
+            messages_dropped=5,
+        )
+
+        assert result.token_count == 100
+        assert result.was_compacted is True
+        assert result.error == "Some error"
+        assert result.original_token_count == 500
+        assert result.messages_summarized == 10
+        assert result.messages_dropped == 5

autogpt_platform/backend/backend/util/request.py

@@ -157,12 +157,7 @@ async def validate_url(
         is_trusted: Boolean indicating if the hostname is in trusted_origins
         ip_addresses: List of IP addresses for the host; empty if the host is trusted
     """
-    # Canonicalize URL
+    parsed = parse_url(url)
-    url = url.strip("/ ").replace("\\", "/")
-    parsed = urlparse(url)
-    if not parsed.scheme:
-        url = f"http://{url}"
-        parsed = urlparse(url)
 
     # Check scheme
     if parsed.scheme not in ALLOWED_SCHEMES:
@@ -220,6 +215,17 @@ async def validate_url(
     )
 
 
+def parse_url(url: str) -> URL:
+    """Canonicalizes and parses a URL string."""
+    url = url.strip("/ ").replace("\\", "/")
+
+    # Ensure scheme is present for proper parsing
+    if not re.match(r"[a-z0-9+.\-]+://", url):
+        url = f"http://{url}"
+
+    return urlparse(url)
+
+
 def pin_url(url: URL, ip_addresses: Optional[list[str]] = None) -> URL:
     """
     Pins a URL to a specific IP address to prevent DNS rebinding attacks.

autogpt_platform/backend/backend/util/settings.py

@@ -656,6 +656,7 @@ class Secrets(UpdateTrackingModel["Secrets"], BaseSettings):
     e2b_api_key: str = Field(default="", description="E2B API key")
     nvidia_api_key: str = Field(default="", description="Nvidia API key")
     mem0_api_key: str = Field(default="", description="Mem0 API key")
+    elevenlabs_api_key: str = Field(default="", description="ElevenLabs API key")
 
     linear_client_id: str = Field(default="", description="Linear client ID")
     linear_client_secret: str = Field(default="", description="Linear client secret")

autogpt_platform/backend/backend/util/workspace.py

@@ -22,6 +22,7 @@ from backend.data.workspace import (
     soft_delete_workspace_file,
 )
 from backend.util.settings import Config
+from backend.util.virus_scanner import scan_content_safe
 from backend.util.workspace_storage import compute_file_checksum, get_workspace_storage
 
 logger = logging.getLogger(__name__)
@@ -187,6 +188,9 @@ class WorkspaceManager:
                 f"{Config().max_file_size_mb}MB limit"
             )
 
+        # Virus scan content before persisting (defense in depth)
+        await scan_content_safe(content, filename=filename)
+
         # Determine path with session scoping
         if path is None:
             path = f"/{filename}"

autogpt_platform/backend/poetry.lock

@@ -825,6 +825,29 @@ files = [
     {file = "charset_normalizer-3.4.2.tar.gz", hash = "sha256:5baececa9ecba31eff645232d59845c07aa030f0c81ee70184a90d35099a0e63"},
 ]
 
+[[package]]
+name = "claude-agent-sdk"
+version = "0.1.31"
+description = "Python SDK for Claude Code"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "claude_agent_sdk-0.1.31-py3-none-macosx_11_0_arm64.whl", hash = "sha256:801bacfe4192782a7cc7b61b0d23a57f061c069993dd3dfa8109aa2e7050a530"},
+    {file = "claude_agent_sdk-0.1.31-py3-none-manylinux_2_17_aarch64.whl", hash = "sha256:0b608e0cbfcedcb827427e6d16a73fe573d58e7f93e15f95435066feacbe6511"},
+    {file = "claude_agent_sdk-0.1.31-py3-none-manylinux_2_17_x86_64.whl", hash = "sha256:d0cb30e026a22246e84d9237d23bb4df20be5146913a04d2802ddd37d4f8b8c9"},
+    {file = "claude_agent_sdk-0.1.31-py3-none-win_amd64.whl", hash = "sha256:8ceca675c2770ad739bd1208362059a830e91c74efcf128045b5a7af14d36f2b"},
+    {file = "claude_agent_sdk-0.1.31.tar.gz", hash = "sha256:b68c681083d7cc985dd3e48f73aabf459f056c1a7e1c5b9c47033c6af94da1a1"},
+]
+
+[package.dependencies]
+anyio = ">=4.0.0"
+mcp = ">=0.1.0"
+typing-extensions = {version = ">=4.0.0", markers = "python_version < \"3.11\""}
+
+[package.extras]
+dev = ["anyio[trio] (>=4.0.0)", "mypy (>=1.0.0)", "pytest (>=7.0.0)", "pytest-asyncio (>=0.20.0)", "pytest-cov (>=4.0.0)", "ruff (>=0.1.0)"]
+
 [[package]]
 name = "cleo"
 version = "2.1.0"
@@ -1169,6 +1192,29 @@ attrs = ">=21.3.0"
 e2b = ">=1.5.4,<2.0.0"
 httpx = ">=0.20.0,<1.0.0"
 
+[[package]]
+name = "elevenlabs"
+version = "1.59.0"
+description = ""
+optional = false
+python-versions = "<4.0,>=3.8"
+groups = ["main"]
+files = [
+    {file = "elevenlabs-1.59.0-py3-none-any.whl", hash = "sha256:468145db81a0bc867708b4a8619699f75583e9481b395ec1339d0b443da771ed"},
+    {file = "elevenlabs-1.59.0.tar.gz", hash = "sha256:16e735bd594e86d415dd445d249c8cc28b09996cfd627fbc10102c0a84698859"},
+]
+
+[package.dependencies]
+httpx = ">=0.21.2"
+pydantic = ">=1.9.2"
+pydantic-core = ">=2.18.2,<3.0.0"
+requests = ">=2.20"
+typing_extensions = ">=4.0.0"
+websockets = ">=11.0"
+
+[package.extras]
+pyaudio = ["pyaudio (>=0.2.14)"]
+
 [[package]]
 name = "email-validator"
 version = "2.2.0"
@@ -2320,6 +2366,18 @@ http2 = ["h2 (>=3,<5)"]
 socks = ["socksio (==1.*)"]
 zstd = ["zstandard (>=0.18.0)"]
 
+[[package]]
+name = "httpx-sse"
+version = "0.4.3"
+description = "Consume Server-Sent Event (SSE) messages with HTTPX."
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "httpx_sse-0.4.3-py3-none-any.whl", hash = "sha256:0ac1c9fe3c0afad2e0ebb25a934a59f4c7823b60792691f779fad2c5568830fc"},
+    {file = "httpx_sse-0.4.3.tar.gz", hash = "sha256:9b1ed0127459a66014aec3c56bebd93da3c1bc8bb6618c8082039a44889a755d"},
+]
+
 [[package]]
 name = "huggingface-hub"
 version = "0.34.4"
@@ -2981,6 +3039,39 @@ files = [
     {file = "mccabe-0.7.0.tar.gz", hash = "sha256:348e0240c33b60bbdf4e523192ef919f28cb2c3d7d5c7794f74009290f236325"},
 ]
 
+[[package]]
+name = "mcp"
+version = "1.26.0"
+description = "Model Context Protocol SDK"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "mcp-1.26.0-py3-none-any.whl", hash = "sha256:904a21c33c25aa98ddbeb47273033c435e595bbacfdb177f4bd87f6dceebe1ca"},
+    {file = "mcp-1.26.0.tar.gz", hash = "sha256:db6e2ef491eecc1a0d93711a76f28dec2e05999f93afd48795da1c1137142c66"},
+]
+
+[package.dependencies]
+anyio = ">=4.5"
+httpx = ">=0.27.1"
+httpx-sse = ">=0.4"
+jsonschema = ">=4.20.0"
+pydantic = ">=2.11.0,<3.0.0"
+pydantic-settings = ">=2.5.2"
+pyjwt = {version = ">=2.10.1", extras = ["crypto"]}
+python-multipart = ">=0.0.9"
+pywin32 = {version = ">=310", markers = "sys_platform == \"win32\""}
+sse-starlette = ">=1.6.1"
+starlette = ">=0.27"
+typing-extensions = ">=4.9.0"
+typing-inspection = ">=0.4.1"
+uvicorn = {version = ">=0.31.1", markers = "sys_platform != \"emscripten\""}
+
+[package.extras]
+cli = ["python-dotenv (>=1.0.0)", "typer (>=0.16.0)"]
+rich = ["rich (>=13.9.4)"]
+ws = ["websockets (>=15.0.1)"]
+
 [[package]]
 name = "mdurl"
 version = "0.1.2"
@@ -5210,7 +5301,7 @@ description = "Python for Window Extensions"
 optional = false
 python-versions = "*"
 groups = ["main"]
-markers = "platform_system == \"Windows\""
+markers = "sys_platform == \"win32\" or platform_system == \"Windows\""
 files = [
     {file = "pywin32-311-cp310-cp310-win32.whl", hash = "sha256:d03ff496d2a0cd4a5893504789d4a15399133fe82517455e78bad62efbb7f0a3"},
     {file = "pywin32-311-cp310-cp310-win_amd64.whl", hash = "sha256:797c2772017851984b97180b0bebe4b620bb86328e8a884bb626156295a63b3b"},
@@ -6195,6 +6286,27 @@ postgresql-psycopgbinary = ["psycopg[binary] (>=3.0.7)"]
 pymysql = ["pymysql"]
 sqlcipher = ["sqlcipher3_binary"]
 
+[[package]]
+name = "sse-starlette"
+version = "3.0.3"
+description = "SSE plugin for Starlette"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "sse_starlette-3.0.3-py3-none-any.whl", hash = "sha256:af5bf5a6f3933df1d9c7f8539633dc8444ca6a97ab2e2a7cd3b6e431ac03a431"},
+    {file = "sse_starlette-3.0.3.tar.gz", hash = "sha256:88cfb08747e16200ea990c8ca876b03910a23b547ab3bd764c0d8eb81019b971"},
+]
+
+[package.dependencies]
+anyio = ">=4.7.0"
+
+[package.extras]
+daphne = ["daphne (>=4.2.0)"]
+examples = ["aiosqlite (>=0.21.0)", "fastapi (>=0.115.12)", "sqlalchemy[asyncio] (>=2.0.41)", "starlette (>=0.49.1)", "uvicorn (>=0.34.0)"]
+granian = ["granian (>=2.3.1)"]
+uvicorn = ["uvicorn (>=0.34.0)"]
+
 [[package]]
 name = "stagehand"
 version = "0.5.1"
@@ -7361,6 +7473,28 @@ files = [
 defusedxml = ">=0.7.1,<0.8.0"
 requests = "*"
 
+[[package]]
+name = "yt-dlp"
+version = "2025.12.8"
+description = "A feature-rich command-line audio/video downloader"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "yt_dlp-2025.12.8-py3-none-any.whl", hash = "sha256:36e2584342e409cfbfa0b5e61448a1c5189e345cf4564294456ee509e7d3e065"},
+    {file = "yt_dlp-2025.12.8.tar.gz", hash = "sha256:b773c81bb6b71cb2c111cfb859f453c7a71cf2ef44eff234ff155877184c3e4f"},
+]
+
+[package.extras]
+build = ["build", "hatchling (>=1.27.0)", "pip", "setuptools (>=71.0.2)", "wheel"]
+curl-cffi = ["curl-cffi (>=0.5.10,<0.6.dev0 || >=0.10.dev0,<0.14) ; implementation_name == \"cpython\""]
+default = ["brotli ; implementation_name == \"cpython\"", "brotlicffi ; implementation_name != \"cpython\"", "certifi", "mutagen", "pycryptodomex", "requests (>=2.32.2,<3)", "urllib3 (>=2.0.2,<3)", "websockets (>=13.0)", "yt-dlp-ejs (==0.3.2)"]
+dev = ["autopep8 (>=2.0,<3.0)", "pre-commit", "pytest (>=8.1,<9.0)", "pytest-rerunfailures (>=14.0,<15.0)", "ruff (>=0.14.0,<0.15.0)"]
+pyinstaller = ["pyinstaller (>=6.17.0)"]
+secretstorage = ["cffi", "secretstorage"]
+static-analysis = ["autopep8 (>=2.0,<3.0)", "ruff (>=0.14.0,<0.15.0)"]
+test = ["pytest (>=8.1,<9.0)", "pytest-rerunfailures (>=14.0,<15.0)"]
+
 [[package]]
 name = "zerobouncesdk"
 version = "1.1.2"
@@ -7512,4 +7646,4 @@ cffi = ["cffi (>=1.11)"]
 [metadata]
 lock-version = "2.1"
 python-versions = ">=3.10,<3.14"
-content-hash = "ee5742dc1a9df50dfc06d4b26a1682cbb2b25cab6b79ce5625ec272f93e4f4bf"
+content-hash = "f79a5f01baf459195d6fd06be2515b83c60cf2aef11a16530842b47febb98a23"

autogpt_platform/backend/pyproject.toml

@@ -13,6 +13,7 @@ aio-pika = "^9.5.5"
 aiohttp = "^3.10.0"
 aiodns = "^3.5.0"
 anthropic = "^0.59.0"
+claude-agent-sdk = "^0.1.0"
 apscheduler = "^3.11.1"
 autogpt-libs = { path = "../autogpt_libs", develop = true }
 bleach = { extras = ["css"], version = "^6.2.0" }
@@ -20,6 +21,7 @@ click = "^8.2.0"
 cryptography = "^45.0"
 discord-py = "^2.5.2"
 e2b-code-interpreter = "^1.5.2"
+elevenlabs = "^1.50.0"
 fastapi = "^0.116.1"
 feedparser = "^6.0.11"
 flake8 = "^7.3.0"
@@ -71,6 +73,7 @@ tweepy = "^4.16.0"
 uvicorn = { extras = ["standard"], version = "^0.35.0" }
 websockets = "^15.0"
 youtube-transcript-api = "^1.2.1"
+yt-dlp = "2025.12.08"
 zerobouncesdk = "^1.1.2"
 # NOTE: please insert new dependencies in their alphabetical location
 pytest-snapshot = "^0.9.0"

autogpt_platform/backend/snapshots/agts_by_creator

@@ -9,7 +9,8 @@
       "sub_heading": "Creator agent subheading",
       "description": "Creator agent description",
       "runs": 50,
-      "rating": 4.0
+      "rating": 4.0,
+      "agent_graph_id": "test-graph-2"
     }
   ],
   "pagination": {

autogpt_platform/backend/snapshots/agts_category

@@ -9,7 +9,8 @@
       "sub_heading": "Category agent subheading",
       "description": "Category agent description",
       "runs": 60,
-      "rating": 4.1
+      "rating": 4.1,
+      "agent_graph_id": "test-graph-category"
     }
   ],
   "pagination": {

autogpt_platform/backend/snapshots/agts_pagination

@@ -9,7 +9,8 @@
       "sub_heading": "Agent 0 subheading",
       "description": "Agent 0 description",
       "runs": 0,
-      "rating": 4.0
+      "rating": 4.0,
+      "agent_graph_id": "test-graph-2"
     },
     {
       "slug": "agent-1",
@@ -20,7 +21,8 @@
       "sub_heading": "Agent 1 subheading",
       "description": "Agent 1 description",
       "runs": 10,
-      "rating": 4.0
+      "rating": 4.0,
+      "agent_graph_id": "test-graph-2"
     },
     {
       "slug": "agent-2",
@@ -31,7 +33,8 @@
       "sub_heading": "Agent 2 subheading",
       "description": "Agent 2 description",
       "runs": 20,
-      "rating": 4.0
+      "rating": 4.0,
+      "agent_graph_id": "test-graph-2"
     },
     {
       "slug": "agent-3",
@@ -42,7 +45,8 @@
       "sub_heading": "Agent 3 subheading",
       "description": "Agent 3 description",
       "runs": 30,
-      "rating": 4.0
+      "rating": 4.0,
+      "agent_graph_id": "test-graph-2"
     },
     {
       "slug": "agent-4",
@@ -53,7 +57,8 @@
       "sub_heading": "Agent 4 subheading",
       "description": "Agent 4 description",
       "runs": 40,
-      "rating": 4.0
+      "rating": 4.0,
+      "agent_graph_id": "test-graph-2"
     }
   ],
   "pagination": {

autogpt_platform/backend/snapshots/agts_search

@@ -9,7 +9,8 @@
       "sub_heading": "Search agent subheading",
       "description": "Specific search term description",
       "runs": 75,
-      "rating": 4.2
+      "rating": 4.2,
+      "agent_graph_id": "test-graph-search"
     }
   ],
   "pagination": {

autogpt_platform/backend/snapshots/agts_sorted

@@ -9,7 +9,8 @@
       "sub_heading": "Top agent subheading",
       "description": "Top agent description",
       "runs": 1000,
-      "rating": 5.0
+      "rating": 5.0,
+      "agent_graph_id": "test-graph-3"
     }
   ],
   "pagination": {

autogpt_platform/backend/snapshots/feat_agts

@@ -9,7 +9,8 @@
       "sub_heading": "Featured agent subheading",
       "description": "Featured agent description",
       "runs": 100,
-      "rating": 4.5
+      "rating": 4.5,
+      "agent_graph_id": "test-graph-1"
     }
   ],
   "pagination": {

autogpt_platform/backend/test/agent_generator/test_core_integration.py

@@ -111,9 +111,7 @@ class TestGenerateAgent:
             instructions = {"type": "instructions", "steps": ["Step 1"]}
             result = await core.generate_agent(instructions)
 
-            # library_agents defaults to None
+            mock_external.assert_called_once_with(instructions, None, None, None)
-            mock_external.assert_called_once_with(instructions, None)
-            # Result should have id, version, is_active added if not present
             assert result is not None
             assert result["name"] == "Test Agent"
             assert "id" in result
@@ -177,8 +175,9 @@ class TestGenerateAgentPatch:
             current_agent = {"nodes": [], "links": []}
             result = await core.generate_agent_patch("Add a node", current_agent)
 
-            # library_agents defaults to None
+            mock_external.assert_called_once_with(
-            mock_external.assert_called_once_with("Add a node", current_agent, None)
+                "Add a node", current_agent, None, None, None
+            )
             assert result == expected_result
 
     @pytest.mark.asyncio

autogpt_platform/backend/test/agent_generator/test_library_agents.py

@@ -134,15 +134,28 @@ class TestSearchMarketplaceAgentsForGeneration:
                 description="A public agent",
                 sub_heading="Does something useful",
                 creator="creator-1",
+                agent_graph_id="graph-123",
             )
         ]
 
-        # The store_db is dynamically imported, so patch the import path
+        mock_graph = MagicMock()
-        with patch(
+        mock_graph.id = "graph-123"
-            "backend.api.features.store.db.get_store_agents",
+        mock_graph.version = 1
-            new_callable=AsyncMock,
+        mock_graph.input_schema = {"type": "object"}
-            return_value=mock_response,
+        mock_graph.output_schema = {"type": "object"}
-        ) as mock_search:
+
+        with (
+            patch(
+                "backend.api.features.store.db.get_store_agents",
+                new_callable=AsyncMock,
+                return_value=mock_response,
+            ) as mock_search,
+            patch(
+                "backend.api.features.chat.tools.agent_generator.core.get_store_listed_graphs",
+                new_callable=AsyncMock,
+                return_value={"graph-123": mock_graph},
+            ),
+        ):
             result = await core.search_marketplace_agents_for_generation(
                 search_query="automation",
                 max_results=10,
@@ -156,7 +169,7 @@ class TestSearchMarketplaceAgentsForGeneration:
 
         assert len(result) == 1
         assert result[0]["name"] == "Public Agent"
-        assert result[0]["is_marketplace_agent"] is True
+        assert result[0]["graph_id"] == "graph-123"
 
     @pytest.mark.asyncio
     async def test_handles_marketplace_error_gracefully(self):
@@ -193,11 +206,12 @@ class TestGetAllRelevantAgentsForGeneration:
 
         marketplace_agents = [
             {
+                "graph_id": "market-456",
+                "graph_version": 1,
                 "name": "Market Agent",
                 "description": "From marketplace",
-                "sub_heading": "Sub heading",
+                "input_schema": {},
-                "creator": "creator-1",
+                "output_schema": {},
-                "is_marketplace_agent": True,
             }
         ]
 
@@ -225,11 +239,11 @@ class TestGetAllRelevantAgentsForGeneration:
         assert result[1]["name"] == "Market Agent"
 
     @pytest.mark.asyncio
-    async def test_deduplicates_by_name(self):
+    async def test_deduplicates_by_graph_id(self):
-        """Test that marketplace agents with same name as library are excluded."""
+        """Test that marketplace agents with same graph_id as library are excluded."""
         library_agents = [
             {
-                "graph_id": "lib-123",
+                "graph_id": "shared-123",
                 "graph_version": 1,
                 "name": "Shared Agent",
                 "description": "From library",
@@ -240,18 +254,20 @@ class TestGetAllRelevantAgentsForGeneration:
 
         marketplace_agents = [
             {
-                "name": "Shared Agent",  # Same name, should be deduplicated
+                "graph_id": "shared-123",  # Same graph_id, should be deduplicated
+                "graph_version": 1,
+                "name": "Shared Agent",
                 "description": "From marketplace",
-                "sub_heading": "Sub heading",
+                "input_schema": {},
-                "creator": "creator-1",
+                "output_schema": {},
-                "is_marketplace_agent": True,
             },
             {
+                "graph_id": "unique-456",
+                "graph_version": 1,
                 "name": "Unique Agent",
                 "description": "Only in marketplace",
-                "sub_heading": "Sub heading",
+                "input_schema": {},
-                "creator": "creator-2",
+                "output_schema": {},
-                "is_marketplace_agent": True,
             },
         ]
 
@@ -273,7 +289,7 @@ class TestGetAllRelevantAgentsForGeneration:
                     include_marketplace=True,
                 )
 
-        # Shared Agent from marketplace should be excluded
+        # Shared Agent from marketplace should be excluded by graph_id
         assert len(result) == 2
         names = [a["name"] for a in result]
         assert "Shared Agent" in names

autogpt_platform/backend/test/agent_generator/test_service.py

@@ -102,7 +102,7 @@ class TestDecomposeGoalExternal:
 
     @pytest.mark.asyncio
     async def test_decompose_goal_with_context(self):
-        """Test decomposition with additional context."""
+        """Test decomposition with additional context enriched into description."""
         mock_response = MagicMock()
         mock_response.json.return_value = {
             "success": True,
@@ -119,9 +119,12 @@ class TestDecomposeGoalExternal:
                 "Build a chatbot", context="Use Python"
             )
 
+        expected_description = (
+            "Build a chatbot\n\nAdditional context from user:\nUse Python"
+        )
         mock_client.post.assert_called_once_with(
             "/api/decompose-description",
-            json={"description": "Build a chatbot", "user_instruction": "Use Python"},
+            json={"description": expected_description},
         )
 
     @pytest.mark.asyncio

autogpt_platform/frontend/src/app/(no-navbar)/onboarding/page.tsx

@@ -1,10 +1,9 @@
 "use client";
+import { getV1OnboardingState } from "@/app/api/__generated__/endpoints/onboarding/onboarding";
+import { getOnboardingStatus, resolveResponse } from "@/app/api/helpers";
 import { LoadingSpinner } from "@/components/atoms/LoadingSpinner/LoadingSpinner";
 import { useRouter } from "next/navigation";
 import { useEffect } from "react";
-import { resolveResponse, getOnboardingStatus } from "@/app/api/helpers";
-import { getV1OnboardingState } from "@/app/api/__generated__/endpoints/onboarding/onboarding";
-import { getHomepageRoute } from "@/lib/constants";
 
 export default function OnboardingPage() {
   const router = useRouter();
@@ -13,12 +12,10 @@ export default function OnboardingPage() {
     async function redirectToStep() {
       try {
         // Check if onboarding is enabled (also gets chat flag for redirect)
-        const { shouldShowOnboarding, isChatEnabled } =
+        const { shouldShowOnboarding } = await getOnboardingStatus();
-          await getOnboardingStatus();
-        const homepageRoute = getHomepageRoute(isChatEnabled);
 
         if (!shouldShowOnboarding) {
-          router.replace(homepageRoute);
+          router.replace("/");
           return;
         }
 
@@ -26,7 +23,7 @@ export default function OnboardingPage() {
 
         // Handle completed onboarding
         if (onboarding.completedSteps.includes("GET_RESULTS")) {
-          router.replace(homepageRoute);
+          router.replace("/");
           return;
         }
 

autogpt_platform/frontend/src/app/(platform)/auth/callback/route.ts

@@ -1,9 +1,8 @@
-import { getServerSupabase } from "@/lib/supabase/server/getServerSupabase";
-import { getHomepageRoute } from "@/lib/constants";
-import BackendAPI from "@/lib/autogpt-server-api";
-import { NextResponse } from "next/server";
-import { revalidatePath } from "next/cache";
 import { getOnboardingStatus } from "@/app/api/helpers";
+import BackendAPI from "@/lib/autogpt-server-api";
+import { getServerSupabase } from "@/lib/supabase/server/getServerSupabase";
+import { revalidatePath } from "next/cache";
+import { NextResponse } from "next/server";
 
 // Handle the callback to complete the user session login
 export async function GET(request: Request) {
@@ -27,13 +26,12 @@ export async function GET(request: Request) {
         await api.createUser();
 
         // Get onboarding status from backend (includes chat flag evaluated for this user)
-        const { shouldShowOnboarding, isChatEnabled } =
+        const { shouldShowOnboarding } = await getOnboardingStatus();
-          await getOnboardingStatus();
         if (shouldShowOnboarding) {
           next = "/onboarding";
           revalidatePath("/onboarding", "layout");
         } else {
-          next = getHomepageRoute(isChatEnabled);
+          next = "/";
           revalidatePath(next, "layout");
         }
       } catch (createUserError) {

autogpt_platform/frontend/src/app/(platform)/auth/integrations/oauth_callback/route.ts

@@ -1,6 +1,17 @@
 import { OAuthPopupResultMessage } from "./types";
 import { NextResponse } from "next/server";
 
+/**
+ * Safely encode a value as JSON for embedding in a script tag.
+ * Escapes characters that could break out of the script context to prevent XSS.
+ */
+function safeJsonStringify(value: unknown): string {
+  return JSON.stringify(value)
+    .replace(/</g, "\\u003c")
+    .replace(/>/g, "\\u003e")
+    .replace(/&/g, "\\u0026");
+}
+
 // This route is intended to be used as the callback for integration OAuth flows,
 // controlled by the CredentialsInput component. The CredentialsInput opens the login
 // page in a pop-up window, which then redirects to this route to close the loop.
@@ -23,12 +34,13 @@ export async function GET(request: Request) {
   console.debug("Sending message to opener:", message);
 
   // Return a response with the message as JSON and a script to close the window
+  // Use safeJsonStringify to prevent XSS by escaping <, >, and & characters
   return new NextResponse(
     `
     <html>
       <body>
         <script>
-          window.opener.postMessage(${JSON.stringify(message)});
+          window.opener.postMessage(${safeJsonStringify(message)});
           window.close();
         </script>
       </body>

autogpt_platform/frontend/src/app/(platform)/build/components/legacy-builder/DataTable.tsx

@@ -1,6 +1,6 @@
 import { beautifyString } from "@/lib/utils";
 import { Clipboard, Maximize2 } from "lucide-react";
-import React, { useState } from "react";
+import React, { useMemo, useState } from "react";
 import { Button } from "../../../../../components/__legacy__/ui/button";
 import { ContentRenderer } from "../../../../../components/__legacy__/ui/render";
 import {
@@ -11,6 +11,12 @@ import {
   TableHeader,
   TableRow,
 } from "../../../../../components/__legacy__/ui/table";
+import type { OutputMetadata } from "@/components/contextual/OutputRenderers";
+import {
+  globalRegistry,
+  OutputItem,
+} from "@/components/contextual/OutputRenderers";
+import { Flag, useGetFlag } from "@/services/feature-flags/use-get-flag";
 import { useToast } from "../../../../../components/molecules/Toast/use-toast";
 import ExpandableOutputDialog from "./ExpandableOutputDialog";
 
@@ -26,6 +32,9 @@ export default function DataTable({
   data,
 }: DataTableProps) {
   const { toast } = useToast();
+  const enableEnhancedOutputHandling = useGetFlag(
+    Flag.ENABLE_ENHANCED_OUTPUT_HANDLING,
+  );
   const [expandedDialog, setExpandedDialog] = useState<{
     isOpen: boolean;
     execId: string;
@@ -33,6 +42,15 @@ export default function DataTable({
     data: any[];
   } | null>(null);
 
+  // Prepare renderers for each item when enhanced mode is enabled
+  const getItemRenderer = useMemo(() => {
+    if (!enableEnhancedOutputHandling) return null;
+    return (item: unknown) => {
+      const metadata: OutputMetadata = {};
+      return globalRegistry.getRenderer(item, metadata);
+    };
+  }, [enableEnhancedOutputHandling]);
+
   const copyData = (pin: string, data: string) => {
     navigator.clipboard.writeText(data).then(() => {
       toast({
@@ -102,15 +120,31 @@ export default function DataTable({
                       <Clipboard size={18} />
                     </Button>
                   </div>
-                  {value.map((item, index) => (
+                  {value.map((item, index) => {
-                    <React.Fragment key={index}>
+                    const renderer = getItemRenderer?.(item);
-                      <ContentRenderer
+                    if (enableEnhancedOutputHandling && renderer) {
-                        value={item}
+                      const metadata: OutputMetadata = {};
-                        truncateLongData={truncateLongData}
+                      return (
-                      />
+                        <React.Fragment key={index}>
-                      {index < value.length - 1 && ", "}
+                          <OutputItem
-                    </React.Fragment>
+                            value={item}
-                  ))}
+                            metadata={metadata}
+                            renderer={renderer}
+                          />
+                          {index < value.length - 1 && ", "}
+                        </React.Fragment>
+                      );
+                    }
+                    return (
+                      <React.Fragment key={index}>
+                        <ContentRenderer
+                          value={item}
+                          truncateLongData={truncateLongData}
+                        />
+                        {index < value.length - 1 && ", "}
+                      </React.Fragment>
+                    );
+                  })}
                 </div>
               </TableCell>
             </TableRow>

autogpt_platform/frontend/src/app/(platform)/build/components/legacy-builder/NodeOutputs.tsx

@@ -1,8 +1,14 @@
-import React, { useContext, useState } from "react";
+import React, { useContext, useMemo, useState } from "react";
 import { Button } from "@/components/__legacy__/ui/button";
 import { Maximize2 } from "lucide-react";
 import * as Separator from "@radix-ui/react-separator";
 import { ContentRenderer } from "@/components/__legacy__/ui/render";
+import type { OutputMetadata } from "@/components/contextual/OutputRenderers";
+import {
+  globalRegistry,
+  OutputItem,
+} from "@/components/contextual/OutputRenderers";
+import { Flag, useGetFlag } from "@/services/feature-flags/use-get-flag";
 
 import { beautifyString } from "@/lib/utils";
 
@@ -21,6 +27,9 @@ export default function NodeOutputs({
   data,
 }: NodeOutputsProps) {
   const builderContext = useContext(BuilderContext);
+  const enableEnhancedOutputHandling = useGetFlag(
+    Flag.ENABLE_ENHANCED_OUTPUT_HANDLING,
+  );
 
   const [expandedDialog, setExpandedDialog] = useState<{
     isOpen: boolean;
@@ -37,6 +46,15 @@ export default function NodeOutputs({
 
   const { getNodeTitle } = builderContext;
 
+  // Prepare renderers for each item when enhanced mode is enabled
+  const getItemRenderer = useMemo(() => {
+    if (!enableEnhancedOutputHandling) return null;
+    return (item: unknown) => {
+      const metadata: OutputMetadata = {};
+      return globalRegistry.getRenderer(item, metadata);
+    };
+  }, [enableEnhancedOutputHandling]);
+
   const getBeautifiedPinName = (pin: string) => {
     if (!pin.startsWith("tools_^_")) {
       return beautifyString(pin);
@@ -87,15 +105,31 @@ export default function NodeOutputs({
           <div className="mt-2">
             <strong className="mr-2">Data:</strong>
             <div className="mt-1">
-              {dataArray.slice(0, 10).map((item, index) => (
+              {dataArray.slice(0, 10).map((item, index) => {
-                <React.Fragment key={index}>
+                const renderer = getItemRenderer?.(item);
-                  <ContentRenderer
+                if (enableEnhancedOutputHandling && renderer) {
-                    value={item}
+                  const metadata: OutputMetadata = {};
-                    truncateLongData={truncateLongData}
+                  return (
-                  />
+                    <React.Fragment key={index}>
-                  {index < Math.min(dataArray.length, 10) - 1 && ", "}
+                      <OutputItem
-                </React.Fragment>
+                        value={item}
-              ))}
+                        metadata={metadata}
+                        renderer={renderer}
+                      />
+                      {index < Math.min(dataArray.length, 10) - 1 && ", "}
+                    </React.Fragment>
+                  );
+                }
+                return (
+                  <React.Fragment key={index}>
+                    <ContentRenderer
+                      value={item}
+                      truncateLongData={truncateLongData}
+                    />
+                    {index < Math.min(dataArray.length, 10) - 1 && ", "}
+                  </React.Fragment>
+                );
+              })}
               {dataArray.length > 10 && (
                 <span style={{ color: "#888" }}>
                   <br />

autogpt_platform/frontend/src/app/(platform)/copilot/components/CopilotShell/useCopilotShell.ts

@@ -11,7 +11,6 @@ import { useBreakpoint } from "@/lib/hooks/useBreakpoint";
 import { useSupabase } from "@/lib/supabase/hooks/useSupabase";
 import { useQueryClient } from "@tanstack/react-query";
 import { usePathname, useSearchParams } from "next/navigation";
-import { useRef } from "react";
 import { useCopilotStore } from "../../copilot-page-store";
 import { useCopilotSessionId } from "../../useCopilotSessionId";
 import { useMobileDrawer } from "./components/MobileDrawer/useMobileDrawer";
@@ -70,41 +69,16 @@ export function useCopilotShell() {
   });
 
   const stopStream = useChatStore((s) => s.stopStream);
-  const onStreamComplete = useChatStore((s) => s.onStreamComplete);
-  const isStreaming = useCopilotStore((s) => s.isStreaming);
   const isCreatingSession = useCopilotStore((s) => s.isCreatingSession);
-  const setIsSwitchingSession = useCopilotStore((s) => s.setIsSwitchingSession);
-  const openInterruptModal = useCopilotStore((s) => s.openInterruptModal);
 
-  const pendingActionRef = useRef<(() => void) | null>(null);
+  function handleSessionClick(sessionId: string) {
-
-  async function stopCurrentStream() {
-    if (!currentSessionId) return;
-
-    setIsSwitchingSession(true);
-    await new Promise<void>((resolve) => {
-      const unsubscribe = onStreamComplete((completedId) => {
-        if (completedId === currentSessionId) {
-          clearTimeout(timeout);
-          unsubscribe();
-          resolve();
-        }
-      });
-      const timeout = setTimeout(() => {
-        unsubscribe();
-        resolve();
-      }, 3000);
-      stopStream(currentSessionId);
-    });
-
-    queryClient.invalidateQueries({
-      queryKey: getGetV2GetSessionQueryKey(currentSessionId),
-    });
-    setIsSwitchingSession(false);
-  }
-
-  function selectSession(sessionId: string) {
     if (sessionId === currentSessionId) return;
+
+    // Stop current stream - SSE reconnection allows resuming later
+    if (currentSessionId) {
+      stopStream(currentSessionId);
+    }
+
     if (recentlyCreatedSessionsRef.current.has(sessionId)) {
       queryClient.invalidateQueries({
         queryKey: getGetV2GetSessionQueryKey(sessionId),
@@ -114,7 +88,12 @@ export function useCopilotShell() {
     if (isMobile) handleCloseDrawer();
   }
 
-  function startNewChat() {
+  function handleNewChatClick() {
+    // Stop current stream - SSE reconnection allows resuming later
+    if (currentSessionId) {
+      stopStream(currentSessionId);
+    }
+
     resetPagination();
     queryClient.invalidateQueries({
       queryKey: getGetV2ListSessionsQueryKey(),
@@ -123,32 +102,6 @@ export function useCopilotShell() {
     if (isMobile) handleCloseDrawer();
   }
 
-  function handleSessionClick(sessionId: string) {
-    if (sessionId === currentSessionId) return;
-
-    if (isStreaming) {
-      pendingActionRef.current = async () => {
-        await stopCurrentStream();
-        selectSession(sessionId);
-      };
-      openInterruptModal(pendingActionRef.current);
-    } else {
-      selectSession(sessionId);
-    }
-  }
-
-  function handleNewChatClick() {
-    if (isStreaming) {
-      pendingActionRef.current = async () => {
-        await stopCurrentStream();
-        startNewChat();
-      };
-      openInterruptModal(pendingActionRef.current);
-    } else {
-      startNewChat();
-    }
-  }
-
   return {
     isMobile,
     isDrawerOpen,

autogpt_platform/frontend/src/app/(platform)/copilot/helpers.ts

@@ -26,8 +26,20 @@ export function buildCopilotChatUrl(prompt: string): string {
 
 export function getQuickActions(): string[] {
   return [
-    "Show me what I can automate",
+    "I don't know where to start, just ask me stuff",
-    "Design a custom workflow",
+    "I do the same thing every week and it's killing me",
-    "Help me with content creation",
+    "Help me find where I'm wasting my time",
   ];
 }
+
+export function getInputPlaceholder(width?: number) {
+  if (!width) return "What's your role and what eats up most of your day?";
+
+  if (width < 500) {
+    return "I'm a chef and I hate...";
+  }
+  if (width <= 1080) {
+    return "What's your role and what eats up most of your day?";
+  }
+  return "What's your role and what eats up most of your day? e.g. 'I'm a recruiter and I hate...'";
+}

autogpt_platform/frontend/src/app/(platform)/copilot/layout.tsx

@@ -1,6 +1,13 @@
-import type { ReactNode } from "react";
+"use client";
+import { FeatureFlagPage } from "@/services/feature-flags/FeatureFlagPage";
+import { Flag } from "@/services/feature-flags/use-get-flag";
+import { type ReactNode } from "react";
 import { CopilotShell } from "./components/CopilotShell/CopilotShell";
 
 export default function CopilotLayout({ children }: { children: ReactNode }) {
-  return <CopilotShell>{children}</CopilotShell>;
+  return (
+    <FeatureFlagPage flag={Flag.CHAT} whenDisabled="/library">
+      <CopilotShell>{children}</CopilotShell>
+    </FeatureFlagPage>
+  );
 }

autogpt_platform/frontend/src/app/(platform)/copilot/page.tsx

@@ -6,7 +6,9 @@ import { Text } from "@/components/atoms/Text/Text";
 import { Chat } from "@/components/contextual/Chat/Chat";
 import { ChatInput } from "@/components/contextual/Chat/components/ChatInput/ChatInput";
 import { Dialog } from "@/components/molecules/Dialog/Dialog";
+import { useEffect, useState } from "react";
 import { useCopilotStore } from "./copilot-page-store";
+import { getInputPlaceholder } from "./helpers";
 import { useCopilotPage } from "./useCopilotPage";
 
 export default function CopilotPage() {
@@ -14,14 +16,25 @@ export default function CopilotPage() {
   const isInterruptModalOpen = useCopilotStore((s) => s.isInterruptModalOpen);
   const confirmInterrupt = useCopilotStore((s) => s.confirmInterrupt);
   const cancelInterrupt = useCopilotStore((s) => s.cancelInterrupt);
-  const {
+
-    greetingName,
+  const [inputPlaceholder, setInputPlaceholder] = useState(
-    quickActions,
+    getInputPlaceholder(),
-    isLoading,
+  );
-    hasSession,
+
-    initialPrompt,
+  useEffect(() => {
-    isReady,
+    const handleResize = () => {
-  } = state;
+      setInputPlaceholder(getInputPlaceholder(window.innerWidth));
+    };
+
+    handleResize();
+
+    window.addEventListener("resize", handleResize);
+    return () => window.removeEventListener("resize", handleResize);
+  }, []);
+
+  const { greetingName, quickActions, isLoading, hasSession, initialPrompt } =
+    state;
+
   const {
     handleQuickAction,
     startChatWithPrompt,
@@ -29,8 +42,6 @@ export default function CopilotPage() {
     handleStreamingChange,
   } = handlers;
 
-  if (!isReady) return null;
-
   if (hasSession) {
     return (
       <div className="flex h-full flex-col">
@@ -81,7 +92,7 @@ export default function CopilotPage() {
   }
 
   return (
-    <div className="flex h-full flex-1 items-center justify-center overflow-y-auto bg-[#f8f8f9] px-6 py-10">
+    <div className="flex h-full flex-1 items-center justify-center overflow-y-auto bg-[#f8f8f9] px-3 py-5 md:px-6 md:py-10">
       <div className="w-full text-center">
         {isLoading ? (
           <div className="mx-auto max-w-2xl">
@@ -98,25 +109,25 @@ export default function CopilotPage() {
           </div>
         ) : (
           <>
-            <div className="mx-auto max-w-2xl">
+            <div className="mx-auto max-w-3xl">
               <Text
                 variant="h3"
-                className="mb-3 !text-[1.375rem] text-zinc-700"
+                className="mb-1 !text-[1.375rem] text-zinc-700"
               >
                 Hey, <span className="text-violet-600">{greetingName}</span>
               </Text>
               <Text variant="h3" className="mb-8 !font-normal">
-                What do you want to automate?
+                Tell me about your work — I&apos;ll find what to automate.
               </Text>
 
               <div className="mb-6">
                 <ChatInput
                   onSend={startChatWithPrompt}
-                  placeholder='You can search or just ask - e.g. "create a blog post outline"'
+                  placeholder={inputPlaceholder}
                 />
               </div>
             </div>
-            <div className="flex flex-nowrap items-center justify-center gap-3 overflow-x-auto [-ms-overflow-style:none] [scrollbar-width:none] [&::-webkit-scrollbar]:hidden">
+            <div className="flex flex-wrap items-center justify-center gap-3 overflow-x-auto [-ms-overflow-style:none] [scrollbar-width:none] [&::-webkit-scrollbar]:hidden">
               {quickActions.map((action) => (
                 <Button
                   key={action}
@@ -124,7 +135,7 @@ export default function CopilotPage() {
                   variant="outline"
                   size="small"
                   onClick={() => handleQuickAction(action)}
-                  className="h-auto shrink-0 border-zinc-600 !px-4 !py-2 text-[1rem] text-zinc-600"
+                  className="h-auto shrink-0 border-zinc-300 px-3 py-2 text-[.9rem] text-zinc-600"
                 >
                   {action}
                 </Button>