fix(chat/sdk): resolve relative paths in security hooks and unify workspace access

The security hook's path validation blocked SDK Read/Write tools because
it didn't resolve relative paths against sdk_cwd. Since the SDK sets cwd,
Claude naturally uses relative paths like "test.txt" which failed the
absolute path prefix check. Now relative paths are joined with sdk_cwd
before validation, and denial messages include the allowed workspace path.

Also clarifies the workspace model: SDK Read/Write + bash_exec share the
same ephemeral session directory, while workspace_file tools provide
persistent cloud storage across sessions.

autogpt_platform/backend/Dockerfile

@@ -62,12 +62,18 @@ ENV POETRY_HOME=/opt/poetry \
     DEBIAN_FRONTEND=noninteractive
 ENV PATH=/opt/poetry/bin:$PATH
 
-# Install Python, FFmpeg, and ImageMagick (required for video processing blocks)
+# Install Python, FFmpeg, ImageMagick, and CLI tools for agent use.
+# bubblewrap provides OS-level sandbox (whitelist-only FS + no network)
+# for the bash_exec MCP tool.
 RUN apt-get update && apt-get install -y \
     python3.13 \
     python3-pip \
     ffmpeg \
     imagemagick \
+    jq \
+    ripgrep \
+    tree \
+    bubblewrap \
     && rm -rf /var/lib/apt/lists/*
 
 # Copy only necessary files from builder

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

@@ -27,12 +27,11 @@ class ChatConfig(BaseSettings):
     session_ttl: int = Field(default=43200, description="Session TTL in seconds")
 
     # Streaming Configuration
-    max_context_messages: int = Field(
-        default=50, ge=1, le=200, description="Maximum context messages"
-    )
-
     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"
@@ -93,6 +92,26 @@ 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",
+    )
+    claude_agent_model: str | None = Field(
+        default=None,
+        description="Model for the Claude Agent SDK path. If None, derives from "
+        "the `model` field by stripping the OpenRouter provider prefix.",
+    )
+    claude_agent_max_buffer_size: int = Field(
+        default=10 * 1024 * 1024,  # 10MB (default SDK is 1MB)
+        description="Max buffer size in bytes for Claude Agent SDK JSON message parsing. "
+        "Increase if tool outputs exceed the limit.",
+    )
+    claude_agent_max_subtasks: int = Field(
+        default=10,
+        description="Max number of sub-agent Tasks the SDK can spawn per session.",
+    )
+
     # Extended thinking configuration for Claude models
     thinking_enabled: bool = Field(
         default=True,
@@ -138,6 +157,17 @@ class ChatConfig(BaseSettings):
             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

@@ -334,9 +334,8 @@ 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:
@@ -378,11 +377,9 @@ 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)
@@ -433,10 +430,9 @@ 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,
@@ -476,7 +472,7 @@ 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:
@@ -493,7 +489,6 @@ 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}")
 
@@ -558,6 +553,40 @@ async def upsert_chat_session(
         return session
 
 
+async def append_and_save_message(session_id: str, message: ChatMessage) -> ChatSession:
+    """Atomically append a message to a session and persist it.
+
+    Acquires the session lock, re-fetches the latest session state,
+    appends the message, and saves — preventing message loss when
+    concurrent requests modify the same session.
+    """
+    lock = await _get_session_lock(session_id)
+
+    async with lock:
+        session = await get_chat_session(session_id)
+        if session is None:
+            raise ValueError(f"Session {session_id} not found")
+
+        session.messages.append(message)
+        existing_message_count = await chat_db.get_chat_session_message_count(
+            session_id
+        )
+
+        try:
+            await _save_session_to_db(session, existing_message_count)
+        except Exception as e:
+            raise DatabaseError(
+                f"Failed to persist message to session {session_id}"
+            ) from e
+
+        try:
+            await _cache_session(session)
+        except Exception as e:
+            logger.warning(f"Cache write failed for session {session_id}: {e}")
+
+        return session
+
+
 async def create_chat_session(user_id: str) -> ChatSession:
     """Create a new chat session and persist it.
 
@@ -664,13 +693,19 @@ 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/routes.py

@@ -1,5 +1,6 @@
 """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
@@ -11,13 +12,22 @@ from fastapi.responses import StreamingResponse
 from pydantic import BaseModel
 
 from backend.util.exceptions import NotFoundError
+from backend.util.feature_flag import Flag, is_feature_enabled
 
 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 (
-from .response_model import StreamFinish, StreamHeartbeat
+    ChatMessage,
+    ChatSession,
+    append_and_save_message,
+    create_chat_session,
+    get_chat_session,
+    get_user_sessions,
+)
+from .response_model import StreamError, StreamFinish, StreamHeartbeat, StreamStart
+from .sdk import service as sdk_service
 from .tools.models import (
     AgentDetailsResponse,
     AgentOutputResponse,
@@ -40,6 +50,7 @@ from .tools.models import (
     SetupRequirementsResponse,
     UnderstandingUpdatedResponse,
 )
+from .tracking import track_user_message
 
 config = ChatConfig()
 
@@ -231,6 +242,10 @@ async def get_session(
     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
@@ -300,10 +315,9 @@ async def stream_chat_post(
         f"user={user_id}, message_len={len(request.message)}",
         extra={"json_fields": log_meta},
     )
-
     session = await _validate_and_get_session(session_id, user_id)
     logger.info(
-        f"[TIMING] session validated in {(time.perf_counter() - stream_start_time)*1000:.1f}ms",
+        f"[TIMING] session validated in {(time.perf_counter() - stream_start_time) * 1000:.1f}ms",
         extra={
             "json_fields": {
                 **log_meta,
@@ -312,6 +326,25 @@ async def stream_chat_post(
         },
     )
 
+    # Atomically append user message to session BEFORE creating task to avoid
+    # race condition where GET_SESSION sees task as "running" but message isn't
+    # saved yet.  append_and_save_message re-fetches inside a lock to prevent
+    # message loss from concurrent requests.
+    if request.message:
+        message = ChatMessage(
+            role="user" if request.is_user_message else "assistant",
+            content=request.message,
+        )
+        if request.is_user_message:
+            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}")
+        session = await append_and_save_message(session_id, message)
+        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())
@@ -327,7 +360,7 @@ async def stream_chat_post(
         operation_id=operation_id,
     )
     logger.info(
-        f"[TIMING] create_task completed in {(time.perf_counter() - task_create_start)*1000:.1f}ms",
+        f"[TIMING] create_task completed in {(time.perf_counter() - task_create_start) * 1000:.1f}ms",
         extra={
             "json_fields": {
                 **log_meta,
@@ -348,15 +381,47 @@ async def stream_chat_post(
         first_chunk_time, ttfc = None, None
         chunk_count = 0
         try:
-            async for chunk in chat_service.stream_chat_completion(
+            # 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)
+            logger.info(
+                f"[TIMING] StreamStart published at {(time_module.perf_counter() - gen_start_time) * 1000:.1f}ms",
+                extra={
+                    "json_fields": {
+                        **log_meta,
+                        "elapsed_ms": (time_module.perf_counter() - gen_start_time)
+                        * 1000,
+                    }
+                },
+            )
+
+            # Choose service based on LaunchDarkly flag (falls back to config default)
+            use_sdk = await is_feature_enabled(
+                Flag.COPILOT_SDK,
+                user_id or "anonymous",
+                default=config.use_claude_agent_sdk,
+            )
+            stream_fn = (
+                sdk_service.stream_chat_completion_sdk
+                if use_sdk
+                else chat_service.stream_chat_completion
+            )
+            logger.info(
+                f"[TIMING] Calling {'sdk' if use_sdk else 'standard'} stream_chat_completion",
+                extra={"json_fields": log_meta},
+            )
+            # Pass message=None since we already added it to the session above
+            async for chunk in stream_fn(
                 session_id,
-                request.message,
+                None,  # Message already in session
                 is_user_message=request.is_user_message,
                 user_id=user_id,
-                session=session,  # Pass pre-fetched session to avoid double-fetch
+                session=session,  # Pass session with message already added
                 context=request.context,
-                _task_id=task_id,  # Pass task_id so service emits start with taskId for reconnection
             ):
+                # Skip duplicate StreamStart — we already published one above
+                if isinstance(chunk, StreamStart):
+                    continue
                 chunk_count += 1
                 if first_chunk_time is None:
                     first_chunk_time = time_module.perf_counter()
@@ -377,7 +442,7 @@ async def stream_chat_post(
             gen_end_time = time_module.perf_counter()
             total_time = (gen_end_time - gen_start_time) * 1000
             logger.info(
-                f"[TIMING] run_ai_generation FINISHED in {total_time/1000:.1f}s; "
+                f"[TIMING] run_ai_generation FINISHED in {total_time / 1000:.1f}s; "
                 f"task={task_id}, session={session_id}, "
                 f"ttfc={ttfc or -1:.2f}s, n_chunks={chunk_count}",
                 extra={
@@ -404,6 +469,17 @@ async def stream_chat_post(
                     }
                 },
             )
+            # Publish a StreamError so the frontend can display an error message
+            try:
+                await stream_registry.publish_chunk(
+                    task_id,
+                    StreamError(
+                        errorText="An error occurred. Please try again.",
+                        code="stream_error",
+                    ),
+                )
+            except Exception:
+                pass  # Best-effort; mark_task_completed will publish StreamFinish
             await stream_registry.mark_task_completed(task_id, "failed")
 
     # Start the AI generation in a background task
@@ -506,8 +582,14 @@ async def stream_chat_post(
                     "json_fields": {**log_meta, "elapsed_ms": elapsed, "error": str(e)}
                 },
             )
+            # Surface error to frontend so it doesn't appear stuck
+            yield StreamError(
+                errorText="An error occurred. Please try again.",
+                code="stream_error",
+            ).to_sse()
+            yield StreamFinish().to_sse()
         finally:
-            # Unsubscribe when client disconnects or stream ends to prevent resource leak
+            # Unsubscribe when client disconnects or stream ends
             if subscriber_queue is not None:
                 try:
                     await stream_registry.unsubscribe_from_task(
@@ -751,8 +833,6 @@ async def stream_task(
         )
 
     async def event_generator() -> AsyncGenerator[str, None]:
-        import asyncio
-
         heartbeat_interval = 15.0  # Send heartbeat every 15 seconds
         try:
             while True:

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/response_adapter.py

@@ -0,0 +1,198 @@
+"""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 claude_agent_sdk import (
+    AssistantMessage,
+    Message,
+    ResultMessage,
+    SystemMessage,
+    TextBlock,
+    ToolResultBlock,
+    ToolUseBlock,
+    UserMessage,
+)
+
+from backend.api.features.chat.response_model import (
+    StreamBaseResponse,
+    StreamError,
+    StreamFinish,
+    StreamFinishStep,
+    StreamStart,
+    StreamStartStep,
+    StreamTextDelta,
+    StreamTextEnd,
+    StreamTextStart,
+    StreamToolInputAvailable,
+    StreamToolInputStart,
+    StreamToolOutputAvailable,
+)
+from backend.api.features.chat.sdk.tool_adapter import (
+    MCP_TOOL_PREFIX,
+    pop_pending_tool_output,
+)
+
+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):
+        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, str]] = {}
+        self.task_id: str | None = None
+        self.step_open = False
+
+    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: Message) -> list[StreamBaseResponse]:
+        """Convert a single SDK message to Vercel AI SDK format."""
+        responses: list[StreamBaseResponse] = []
+
+        if isinstance(sdk_message, SystemMessage):
+            if sdk_message.subtype == "init":
+                responses.append(
+                    StreamStart(messageId=self.message_id, taskId=self.task_id)
+                )
+                # Open the first step (matches non-SDK: StreamStart then StreamStartStep)
+                responses.append(StreamStartStep())
+                self.step_open = True
+
+        elif isinstance(sdk_message, AssistantMessage):
+            # After tool results, the SDK sends a new AssistantMessage for the
+            # next LLM turn. Open a new step if the previous one was closed.
+            if not self.step_open:
+                responses.append(StreamStartStep())
+                self.step_open = True
+
+            for block in sdk_message.content:
+                if isinstance(block, TextBlock):
+                    if block.text:
+                        self._ensure_text_started(responses)
+                        responses.append(
+                            StreamTextDelta(id=self.text_block_id, delta=block.text)
+                        )
+
+                elif isinstance(block, ToolUseBlock):
+                    self._end_text_if_open(responses)
+
+                    # Strip MCP prefix so frontend sees "find_block"
+                    # instead of "mcp__copilot__find_block".
+                    tool_name = block.name.removeprefix(MCP_TOOL_PREFIX)
+
+                    responses.append(
+                        StreamToolInputStart(toolCallId=block.id, toolName=tool_name)
+                    )
+                    responses.append(
+                        StreamToolInputAvailable(
+                            toolCallId=block.id,
+                            toolName=tool_name,
+                            input=block.input,
+                        )
+                    )
+                    self.current_tool_calls[block.id] = {"name": tool_name}
+
+        elif isinstance(sdk_message, UserMessage):
+            # UserMessage carries tool results back from tool execution.
+            content = sdk_message.content
+            blocks = content if isinstance(content, list) else []
+            for block in blocks:
+                if isinstance(block, ToolResultBlock) and block.tool_use_id:
+                    tool_info = self.current_tool_calls.get(block.tool_use_id, {})
+                    tool_name = tool_info.get("name", "unknown")
+
+                    # Prefer the stashed full output over the SDK's
+                    # (potentially truncated) ToolResultBlock content.
+                    # The SDK truncates large results, writing them to disk,
+                    # which breaks frontend widget parsing.
+                    output = pop_pending_tool_output(tool_name) or (
+                        _extract_tool_output(block.content)
+                    )
+
+                    responses.append(
+                        StreamToolOutputAvailable(
+                            toolCallId=block.tool_use_id,
+                            toolName=tool_name,
+                            output=output,
+                            success=not (block.is_error or False),
+                        )
+                    )
+
+            # Close the current step after tool results — the next
+            # AssistantMessage will open a new step for the continuation.
+            if self.step_open:
+                responses.append(StreamFinishStep())
+                self.step_open = False
+
+        elif isinstance(sdk_message, ResultMessage):
+            self._end_text_if_open(responses)
+            # Close the step before finishing.
+            if self.step_open:
+                responses.append(StreamFinishStep())
+                self.step_open = False
+
+            if sdk_message.subtype == "success":
+                responses.append(StreamFinish())
+            elif sdk_message.subtype in ("error", "error_during_execution"):
+                error_msg = getattr(sdk_message, "result", None) or "Unknown error"
+                responses.append(
+                    StreamError(errorText=str(error_msg), code="sdk_error")
+                )
+                responses.append(StreamFinish())
+
+        else:
+            logger.debug(f"Unhandled SDK message type: {type(sdk_message).__name__}")
+
+        return responses
+
+    def _ensure_text_started(self, responses: list[StreamBaseResponse]) -> None:
+        """Start (or restart) a text block if needed."""
+        if not self.has_started_text or self.has_ended_text:
+            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
+
+    def _end_text_if_open(self, responses: list[StreamBaseResponse]) -> None:
+        """End the current text block if one is 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
+
+
+def _extract_tool_output(content: str | list[dict[str, str]] | None) -> str:
+    """Extract a string output from a ToolResultBlock's content field."""
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        parts = [item.get("text", "") for item in content if item.get("type") == "text"]
+        if parts:
+            return "".join(parts)
+        try:
+            return json.dumps(content)
+        except (TypeError, ValueError):
+            return str(content)
+    if content is None:
+        return ""
+    try:
+        return json.dumps(content)
+    except (TypeError, ValueError):
+        return str(content)

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

@@ -0,0 +1,366 @@
+"""Unit tests for the SDK response adapter."""
+
+from claude_agent_sdk import (
+    AssistantMessage,
+    ResultMessage,
+    SystemMessage,
+    TextBlock,
+    ToolResultBlock,
+    ToolUseBlock,
+    UserMessage,
+)
+
+from backend.api.features.chat.response_model import (
+    StreamBaseResponse,
+    StreamError,
+    StreamFinish,
+    StreamFinishStep,
+    StreamStart,
+    StreamStartStep,
+    StreamTextDelta,
+    StreamTextEnd,
+    StreamTextStart,
+    StreamToolInputAvailable,
+    StreamToolInputStart,
+    StreamToolOutputAvailable,
+)
+
+from .response_adapter import SDKResponseAdapter
+from .tool_adapter import MCP_TOOL_PREFIX
+
+
+def _adapter() -> SDKResponseAdapter:
+    a = SDKResponseAdapter(message_id="msg-1")
+    a.set_task_id("task-1")
+    return a
+
+
+# -- SystemMessage -----------------------------------------------------------
+
+
+def test_system_init_emits_start_and_step():
+    adapter = _adapter()
+    results = adapter.convert_message(SystemMessage(subtype="init", data={}))
+    assert len(results) == 2
+    assert isinstance(results[0], StreamStart)
+    assert results[0].messageId == "msg-1"
+    assert results[0].taskId == "task-1"
+    assert isinstance(results[1], StreamStartStep)
+
+
+def test_system_non_init_emits_nothing():
+    adapter = _adapter()
+    results = adapter.convert_message(SystemMessage(subtype="other", data={}))
+    assert results == []
+
+
+# -- AssistantMessage with TextBlock -----------------------------------------
+
+
+def test_text_block_emits_step_start_and_delta():
+    adapter = _adapter()
+    msg = AssistantMessage(content=[TextBlock(text="hello")], model="test")
+    results = adapter.convert_message(msg)
+    assert len(results) == 3
+    assert isinstance(results[0], StreamStartStep)
+    assert isinstance(results[1], StreamTextStart)
+    assert isinstance(results[2], StreamTextDelta)
+    assert results[2].delta == "hello"
+
+
+def test_empty_text_block_emits_only_step():
+    adapter = _adapter()
+    msg = AssistantMessage(content=[TextBlock(text="")], model="test")
+    results = adapter.convert_message(msg)
+    # Empty text skipped, but step still opens
+    assert len(results) == 1
+    assert isinstance(results[0], StreamStartStep)
+
+
+def test_multiple_text_deltas_reuse_block_id():
+    adapter = _adapter()
+    msg1 = AssistantMessage(content=[TextBlock(text="a")], model="test")
+    msg2 = AssistantMessage(content=[TextBlock(text="b")], model="test")
+    r1 = adapter.convert_message(msg1)
+    r2 = adapter.convert_message(msg2)
+    # First gets step+start+delta, second only delta (block & step already started)
+    assert len(r1) == 3
+    assert isinstance(r1[0], StreamStartStep)
+    assert isinstance(r1[1], StreamTextStart)
+    assert len(r2) == 1
+    assert isinstance(r2[0], StreamTextDelta)
+    assert r1[1].id == r2[0].id  # same block ID
+
+
+# -- AssistantMessage with ToolUseBlock --------------------------------------
+
+
+def test_tool_use_emits_input_start_and_available():
+    """Tool names arrive with MCP prefix and should be stripped for the frontend."""
+    adapter = _adapter()
+    msg = AssistantMessage(
+        content=[
+            ToolUseBlock(
+                id="tool-1",
+                name=f"{MCP_TOOL_PREFIX}find_agent",
+                input={"q": "x"},
+            )
+        ],
+        model="test",
+    )
+    results = adapter.convert_message(msg)
+    assert len(results) == 3
+    assert isinstance(results[0], StreamStartStep)
+    assert isinstance(results[1], StreamToolInputStart)
+    assert results[1].toolCallId == "tool-1"
+    assert results[1].toolName == "find_agent"  # prefix stripped
+    assert isinstance(results[2], StreamToolInputAvailable)
+    assert results[2].toolName == "find_agent"  # prefix stripped
+    assert results[2].input == {"q": "x"}
+
+
+def test_text_then_tool_ends_text_block():
+    adapter = _adapter()
+    text_msg = AssistantMessage(content=[TextBlock(text="thinking...")], model="test")
+    tool_msg = AssistantMessage(
+        content=[ToolUseBlock(id="t1", name=f"{MCP_TOOL_PREFIX}tool", input={})],
+        model="test",
+    )
+    adapter.convert_message(text_msg)  # opens step + text
+    results = adapter.convert_message(tool_msg)
+    # Step already open, so: TextEnd, ToolInputStart, ToolInputAvailable
+    assert len(results) == 3
+    assert isinstance(results[0], StreamTextEnd)
+    assert isinstance(results[1], StreamToolInputStart)
+
+
+# -- UserMessage with ToolResultBlock ----------------------------------------
+
+
+def test_tool_result_emits_output_and_finish_step():
+    adapter = _adapter()
+    # First register the tool call (opens step) — SDK sends prefixed name
+    tool_msg = AssistantMessage(
+        content=[ToolUseBlock(id="t1", name=f"{MCP_TOOL_PREFIX}find_agent", input={})],
+        model="test",
+    )
+    adapter.convert_message(tool_msg)
+
+    # Now send tool result
+    result_msg = UserMessage(
+        content=[ToolResultBlock(tool_use_id="t1", content="found 3 agents")]
+    )
+    results = adapter.convert_message(result_msg)
+    assert len(results) == 2
+    assert isinstance(results[0], StreamToolOutputAvailable)
+    assert results[0].toolCallId == "t1"
+    assert results[0].toolName == "find_agent"  # prefix stripped
+    assert results[0].output == "found 3 agents"
+    assert results[0].success is True
+    assert isinstance(results[1], StreamFinishStep)
+
+
+def test_tool_result_error():
+    adapter = _adapter()
+    adapter.convert_message(
+        AssistantMessage(
+            content=[
+                ToolUseBlock(id="t1", name=f"{MCP_TOOL_PREFIX}run_agent", input={})
+            ],
+            model="test",
+        )
+    )
+    result_msg = UserMessage(
+        content=[ToolResultBlock(tool_use_id="t1", content="timeout", is_error=True)]
+    )
+    results = adapter.convert_message(result_msg)
+    assert isinstance(results[0], StreamToolOutputAvailable)
+    assert results[0].success is False
+    assert isinstance(results[1], StreamFinishStep)
+
+
+def test_tool_result_list_content():
+    adapter = _adapter()
+    adapter.convert_message(
+        AssistantMessage(
+            content=[ToolUseBlock(id="t1", name=f"{MCP_TOOL_PREFIX}tool", input={})],
+            model="test",
+        )
+    )
+    result_msg = UserMessage(
+        content=[
+            ToolResultBlock(
+                tool_use_id="t1",
+                content=[
+                    {"type": "text", "text": "line1"},
+                    {"type": "text", "text": "line2"},
+                ],
+            )
+        ]
+    )
+    results = adapter.convert_message(result_msg)
+    assert isinstance(results[0], StreamToolOutputAvailable)
+    assert results[0].output == "line1line2"
+    assert isinstance(results[1], StreamFinishStep)
+
+
+def test_string_user_message_ignored():
+    """A plain string UserMessage (not tool results) produces no output."""
+    adapter = _adapter()
+    results = adapter.convert_message(UserMessage(content="hello"))
+    assert results == []
+
+
+# -- ResultMessage -----------------------------------------------------------
+
+
+def test_result_success_emits_finish_step_and_finish():
+    adapter = _adapter()
+    # Start some text first (opens step)
+    adapter.convert_message(
+        AssistantMessage(content=[TextBlock(text="done")], model="test")
+    )
+    msg = ResultMessage(
+        subtype="success",
+        duration_ms=100,
+        duration_api_ms=50,
+        is_error=False,
+        num_turns=1,
+        session_id="s1",
+    )
+    results = adapter.convert_message(msg)
+    # TextEnd + FinishStep + StreamFinish
+    assert len(results) == 3
+    assert isinstance(results[0], StreamTextEnd)
+    assert isinstance(results[1], StreamFinishStep)
+    assert isinstance(results[2], StreamFinish)
+
+
+def test_result_error_emits_error_and_finish():
+    adapter = _adapter()
+    msg = ResultMessage(
+        subtype="error",
+        duration_ms=100,
+        duration_api_ms=50,
+        is_error=True,
+        num_turns=0,
+        session_id="s1",
+        result="API rate limited",
+    )
+    results = adapter.convert_message(msg)
+    # No step was open, so no FinishStep — just Error + Finish
+    assert len(results) == 2
+    assert isinstance(results[0], StreamError)
+    assert "API rate limited" in results[0].errorText
+    assert isinstance(results[1], StreamFinish)
+
+
+# -- Text after tools (new block ID) ----------------------------------------
+
+
+def test_text_after_tool_gets_new_block_id():
+    adapter = _adapter()
+    # Text -> Tool -> ToolResult -> Text should get a new text block ID and step
+    adapter.convert_message(
+        AssistantMessage(content=[TextBlock(text="before")], model="test")
+    )
+    adapter.convert_message(
+        AssistantMessage(
+            content=[ToolUseBlock(id="t1", name=f"{MCP_TOOL_PREFIX}tool", input={})],
+            model="test",
+        )
+    )
+    # Send tool result (closes step)
+    adapter.convert_message(
+        UserMessage(content=[ToolResultBlock(tool_use_id="t1", content="ok")])
+    )
+    results = adapter.convert_message(
+        AssistantMessage(content=[TextBlock(text="after")], model="test")
+    )
+    # Should get StreamStartStep (new step) + StreamTextStart (new block) + StreamTextDelta
+    assert len(results) == 3
+    assert isinstance(results[0], StreamStartStep)
+    assert isinstance(results[1], StreamTextStart)
+    assert isinstance(results[2], StreamTextDelta)
+    assert results[2].delta == "after"
+
+
+# -- Full conversation flow --------------------------------------------------
+
+
+def test_full_conversation_flow():
+    """Simulate a complete conversation: init -> text -> tool -> result -> text -> finish."""
+    adapter = _adapter()
+    all_responses: list[StreamBaseResponse] = []
+
+    # 1. Init
+    all_responses.extend(
+        adapter.convert_message(SystemMessage(subtype="init", data={}))
+    )
+    # 2. Assistant text
+    all_responses.extend(
+        adapter.convert_message(
+            AssistantMessage(content=[TextBlock(text="Let me search")], model="test")
+        )
+    )
+    # 3. Tool use
+    all_responses.extend(
+        adapter.convert_message(
+            AssistantMessage(
+                content=[
+                    ToolUseBlock(
+                        id="t1",
+                        name=f"{MCP_TOOL_PREFIX}find_agent",
+                        input={"query": "email"},
+                    )
+                ],
+                model="test",
+            )
+        )
+    )
+    # 4. Tool result
+    all_responses.extend(
+        adapter.convert_message(
+            UserMessage(
+                content=[ToolResultBlock(tool_use_id="t1", content="Found 2 agents")]
+            )
+        )
+    )
+    # 5. More text
+    all_responses.extend(
+        adapter.convert_message(
+            AssistantMessage(content=[TextBlock(text="I found 2")], model="test")
+        )
+    )
+    # 6. Result
+    all_responses.extend(
+        adapter.convert_message(
+            ResultMessage(
+                subtype="success",
+                duration_ms=500,
+                duration_api_ms=400,
+                is_error=False,
+                num_turns=2,
+                session_id="s1",
+            )
+        )
+    )
+
+    types = [type(r).__name__ for r in all_responses]
+    assert types == [
+        "StreamStart",
+        "StreamStartStep",  # step 1: text + tool call
+        "StreamTextStart",
+        "StreamTextDelta",  # "Let me search"
+        "StreamTextEnd",  # closed before tool
+        "StreamToolInputStart",
+        "StreamToolInputAvailable",
+        "StreamToolOutputAvailable",  # tool result
+        "StreamFinishStep",  # step 1 closed after tool result
+        "StreamStartStep",  # step 2: continuation text
+        "StreamTextStart",  # new block after tool
+        "StreamTextDelta",  # "I found 2"
+        "StreamTextEnd",  # closed by result
+        "StreamFinishStep",  # step 2 closed
+        "StreamFinish",
+    ]

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

@@ -0,0 +1,296 @@
+"""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 json
+import logging
+import os
+import re
+from typing import Any, cast
+
+from backend.api.features.chat.sdk.tool_adapter import MCP_TOOL_PREFIX
+
+logger = logging.getLogger(__name__)
+
+# Tools that are blocked entirely (CLI/system access).
+# "Bash" (capital) is the SDK built-in — it's NOT in allowed_tools but blocked
+# here as defence-in-depth.  The agent uses mcp__copilot__bash_exec instead,
+# which has kernel-level network isolation (unshare --net).
+BLOCKED_TOOLS = {
+    "Bash",
+    "bash",
+    "shell",
+    "exec",
+    "terminal",
+    "command",
+}
+
+# Tools allowed only when their path argument stays within the SDK workspace.
+# The SDK uses these to handle oversized tool results (writes to tool-results/
+# files, then reads them back) and for workspace file operations.
+WORKSPACE_SCOPED_TOOLS = {"Read", "Write", "Edit", "Glob", "Grep"}
+
+# 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 _deny(reason: str) -> dict[str, Any]:
+    """Return a hook denial response."""
+    return {
+        "hookSpecificOutput": {
+            "hookEventName": "PreToolUse",
+            "permissionDecision": "deny",
+            "permissionDecisionReason": reason,
+        }
+    }
+
+
+def _validate_workspace_path(
+    tool_name: str, tool_input: dict[str, Any], sdk_cwd: str | None
+) -> dict[str, Any]:
+    """Validate that a workspace-scoped tool only accesses allowed paths.
+
+    Allowed directories:
+    - The SDK working directory (``/tmp/copilot-<session>/``)
+    - The SDK tool-results directory (``~/.claude/projects/…/tool-results/``)
+    """
+    path = tool_input.get("file_path") or tool_input.get("path") or ""
+    if not path:
+        # Glob/Grep without a path default to cwd which is already sandboxed
+        return {}
+
+    # Resolve relative paths against sdk_cwd (the SDK sets cwd so the LLM
+    # naturally uses relative paths like "test.txt" instead of absolute ones).
+    if not os.path.isabs(path) and sdk_cwd:
+        resolved = os.path.normpath(os.path.join(sdk_cwd, path))
+    else:
+        resolved = os.path.normpath(os.path.expanduser(path))
+
+    # Allow access within the SDK working directory
+    if sdk_cwd:
+        norm_cwd = os.path.normpath(sdk_cwd)
+        if resolved.startswith(norm_cwd + os.sep) or resolved == norm_cwd:
+            return {}
+
+    # Allow access to ~/.claude/projects/*/tool-results/ (big tool results)
+    claude_dir = os.path.normpath(os.path.expanduser("~/.claude/projects"))
+    if resolved.startswith(claude_dir + os.sep) and "tool-results" in resolved:
+        return {}
+
+    logger.warning(
+        f"Blocked {tool_name} outside workspace: {path} (resolved={resolved})"
+    )
+    workspace_hint = f" Allowed workspace: {sdk_cwd}" if sdk_cwd else ""
+    return _deny(
+        f"[SECURITY] Tool '{tool_name}' can only access files within the workspace "
+        f"directory.{workspace_hint} "
+        "This is enforced by the platform and cannot be bypassed."
+    )
+
+
+def _validate_tool_access(
+    tool_name: str, tool_input: dict[str, Any], sdk_cwd: str | None = None
+) -> 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 _deny(
+            f"[SECURITY] Tool '{tool_name}' is blocked for security. "
+            "This is enforced by the platform and cannot be bypassed. "
+            "Use the CoPilot-specific MCP tools instead."
+        )
+
+    # Workspace-scoped tools: allowed only within the SDK workspace directory
+    if tool_name in WORKSPACE_SCOPED_TOOLS:
+        return _validate_workspace_path(tool_name, tool_input, sdk_cwd)
+
+    # Check for dangerous patterns in tool input
+    # Use json.dumps for predictable format (str() produces Python repr)
+    input_str = json.dumps(tool_input) if tool_input else ""
+
+    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 _deny(
+                "[SECURITY] Input contains a blocked pattern. "
+                "This is enforced by the platform and cannot be bypassed."
+            )
+
+    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,
+    sdk_cwd: str | None = None,
+    max_subtasks: int = 3,
+) -> 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
+        sdk_cwd: SDK working directory for workspace-scoped tool validation
+        max_subtasks: Maximum Task (sub-agent) spawns allowed per session
+
+    Returns:
+        Hooks configuration dict for ClaudeAgentOptions
+    """
+    try:
+        from claude_agent_sdk import HookMatcher
+        from claude_agent_sdk.types import HookContext, HookInput, SyncHookJSONOutput
+
+        # Per-session counter for Task sub-agent spawns
+        task_spawn_count = 0
+
+        async def pre_tool_use_hook(
+            input_data: HookInput,
+            tool_use_id: str | None,
+            context: HookContext,
+        ) -> SyncHookJSONOutput:
+            """Combined pre-tool-use validation hook."""
+            nonlocal task_spawn_count
+            _ = 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", {}))
+
+            # Rate-limit Task (sub-agent) spawns per session
+            if tool_name == "Task":
+                task_spawn_count += 1
+                if task_spawn_count > max_subtasks:
+                    logger.warning(
+                        f"[SDK] Task limit reached ({max_subtasks}), user={user_id}"
+                    )
+                    return cast(
+                        SyncHookJSONOutput,
+                        _deny(
+                            f"Maximum {max_subtasks} sub-tasks per session. "
+                            "Please continue in the main conversation."
+                        ),
+                    )
+
+            # Strip MCP prefix for consistent validation
+            is_copilot_tool = tool_name.startswith(MCP_TOOL_PREFIX)
+            clean_name = tool_name.removeprefix(MCP_TOOL_PREFIX)
+
+            # Only block non-CoPilot tools; our MCP-registered tools
+            # (including Read for oversized results) are already sandboxed.
+            if not is_copilot_tool:
+                result = _validate_tool_access(clean_name, tool_input, sdk_cwd)
+                if result:
+                    return cast(SyncHookJSONOutput, result)
+
+            # Validate user isolation
+            result = _validate_user_isolation(clean_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
+        logger.warning("claude-agent-sdk not available, security hooks disabled")
+        return {}

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

@@ -0,0 +1,165 @@
+"""Unit tests for SDK security hooks."""
+
+import os
+
+from .security_hooks import _validate_tool_access, _validate_user_isolation
+
+SDK_CWD = "/tmp/copilot-abc123"
+
+
+def _is_denied(result: dict) -> bool:
+    hook = result.get("hookSpecificOutput", {})
+    return hook.get("permissionDecision") == "deny"
+
+
+# -- Blocked tools -----------------------------------------------------------
+
+
+def test_blocked_tools_denied():
+    for tool in ("bash", "shell", "exec", "terminal", "command"):
+        result = _validate_tool_access(tool, {})
+        assert _is_denied(result), f"{tool} should be blocked"
+
+
+def test_unknown_tool_allowed():
+    result = _validate_tool_access("SomeCustomTool", {})
+    assert result == {}
+
+
+# -- Workspace-scoped tools --------------------------------------------------
+
+
+def test_read_within_workspace_allowed():
+    result = _validate_tool_access(
+        "Read", {"file_path": f"{SDK_CWD}/file.txt"}, sdk_cwd=SDK_CWD
+    )
+    assert result == {}
+
+
+def test_write_within_workspace_allowed():
+    result = _validate_tool_access(
+        "Write", {"file_path": f"{SDK_CWD}/output.json"}, sdk_cwd=SDK_CWD
+    )
+    assert result == {}
+
+
+def test_edit_within_workspace_allowed():
+    result = _validate_tool_access(
+        "Edit", {"file_path": f"{SDK_CWD}/src/main.py"}, sdk_cwd=SDK_CWD
+    )
+    assert result == {}
+
+
+def test_glob_within_workspace_allowed():
+    result = _validate_tool_access("Glob", {"path": f"{SDK_CWD}/src"}, sdk_cwd=SDK_CWD)
+    assert result == {}
+
+
+def test_grep_within_workspace_allowed():
+    result = _validate_tool_access("Grep", {"path": f"{SDK_CWD}/src"}, sdk_cwd=SDK_CWD)
+    assert result == {}
+
+
+def test_read_outside_workspace_denied():
+    result = _validate_tool_access(
+        "Read", {"file_path": "/etc/passwd"}, sdk_cwd=SDK_CWD
+    )
+    assert _is_denied(result)
+
+
+def test_write_outside_workspace_denied():
+    result = _validate_tool_access(
+        "Write", {"file_path": "/home/user/secrets.txt"}, sdk_cwd=SDK_CWD
+    )
+    assert _is_denied(result)
+
+
+def test_traversal_attack_denied():
+    result = _validate_tool_access(
+        "Read",
+        {"file_path": f"{SDK_CWD}/../../etc/passwd"},
+        sdk_cwd=SDK_CWD,
+    )
+    assert _is_denied(result)
+
+
+def test_no_path_allowed():
+    """Glob/Grep without a path argument defaults to cwd — should pass."""
+    result = _validate_tool_access("Glob", {}, sdk_cwd=SDK_CWD)
+    assert result == {}
+
+
+def test_read_no_cwd_denies_absolute():
+    """If no sdk_cwd is set, absolute paths are denied."""
+    result = _validate_tool_access("Read", {"file_path": "/tmp/anything"})
+    assert _is_denied(result)
+
+
+# -- Tool-results directory --------------------------------------------------
+
+
+def test_read_tool_results_allowed():
+    home = os.path.expanduser("~")
+    path = f"{home}/.claude/projects/-tmp-copilot-abc123/tool-results/12345.txt"
+    result = _validate_tool_access("Read", {"file_path": path}, sdk_cwd=SDK_CWD)
+    assert result == {}
+
+
+def test_read_claude_projects_without_tool_results_denied():
+    home = os.path.expanduser("~")
+    path = f"{home}/.claude/projects/-tmp-copilot-abc123/settings.json"
+    result = _validate_tool_access("Read", {"file_path": path}, sdk_cwd=SDK_CWD)
+    assert _is_denied(result)
+
+
+# -- Built-in Bash is blocked (use bash_exec MCP tool instead) ---------------
+
+
+def test_bash_builtin_always_blocked():
+    """SDK built-in Bash is blocked — bash_exec MCP tool with bubblewrap is used instead."""
+    result = _validate_tool_access("Bash", {"command": "echo hello"}, sdk_cwd=SDK_CWD)
+    assert _is_denied(result)
+
+
+# -- Dangerous patterns ------------------------------------------------------
+
+
+def test_dangerous_pattern_blocked():
+    result = _validate_tool_access("SomeTool", {"cmd": "sudo rm -rf /"})
+    assert _is_denied(result)
+
+
+def test_subprocess_pattern_blocked():
+    result = _validate_tool_access("SomeTool", {"code": "subprocess.run(...)"})
+    assert _is_denied(result)
+
+
+# -- User isolation ----------------------------------------------------------
+
+
+def test_workspace_path_traversal_blocked():
+    result = _validate_user_isolation(
+        "workspace_read", {"path": "../../../etc/shadow"}, user_id="user-1"
+    )
+    assert _is_denied(result)
+
+
+def test_workspace_absolute_path_blocked():
+    result = _validate_user_isolation(
+        "workspace_read", {"path": "/etc/passwd"}, user_id="user-1"
+    )
+    assert _is_denied(result)
+
+
+def test_workspace_normal_path_allowed():
+    result = _validate_user_isolation(
+        "workspace_read", {"path": "src/main.py"}, user_id="user-1"
+    )
+    assert result == {}
+
+
+def test_non_workspace_tool_passes_isolation():
+    result = _validate_user_isolation(
+        "find_agent", {"query": "email"}, user_id="user-1"
+    )
+    assert result == {}

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

@@ -0,0 +1,668 @@
+"""Claude Agent SDK service layer for CoPilot chat completions."""
+
+import asyncio
+import json
+import logging
+import os
+import uuid
+from collections.abc import AsyncGenerator
+from typing import Any
+
+from backend.util.exceptions import NotFoundError
+
+from .. import stream_registry
+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 ..service import (
+    _build_system_prompt,
+    _execute_long_running_tool_with_streaming,
+    _generate_session_title,
+)
+from ..tools.models import OperationPendingResponse, OperationStartedResponse
+from ..tools.sandbox import WORKSPACE_PREFIX, make_session_path
+from ..tracking import track_user_message
+from .response_adapter import SDKResponseAdapter
+from .security_hooks import create_security_hooks
+from .tool_adapter import (
+    COPILOT_TOOL_NAMES,
+    LongRunningCallback,
+    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()
+
+
+_SDK_CWD_PREFIX = WORKSPACE_PREFIX
+
+# Appended to the system prompt to inform the agent about available tools.
+# The SDK built-in Bash is NOT available — use mcp__copilot__bash_exec instead,
+# which has kernel-level network isolation (unshare --net).
+_SDK_TOOL_SUPPLEMENT = """
+
+## Tool notes
+
+- The SDK built-in Bash tool is NOT available.  Use the `bash_exec` MCP tool
+  for shell commands — it runs in a network-isolated sandbox.
+- **Shared workspace**: The SDK Read/Write tools and `bash_exec` share the
+  same working directory. Files created by one are readable by the other.
+  These files are **ephemeral** — they exist only for the current session.
+- **Persistent storage**: Use `write_workspace_file` / `read_workspace_file`
+  for files that should persist across sessions (stored in cloud storage).
+- Long-running tools (create_agent, edit_agent, etc.) are handled
+  asynchronously.  You will receive an immediate response; the actual result
+  is delivered to the user via a background stream.
+"""
+
+
+def _build_long_running_callback(user_id: str | None) -> LongRunningCallback:
+    """Build a callback that delegates long-running tools to the non-SDK infrastructure.
+
+    Long-running tools (create_agent, edit_agent, etc.) are delegated to the
+    existing background infrastructure: stream_registry (Redis Streams),
+    database persistence, and SSE reconnection.  This means results survive
+    page refreshes / pod restarts, and the frontend shows the proper loading
+    widget with progress updates.
+
+    The returned callback matches the ``LongRunningCallback`` signature:
+    ``(tool_name, args, session) -> MCP response dict``.
+    """
+
+    async def _callback(
+        tool_name: str, args: dict[str, Any], session: ChatSession
+    ) -> dict[str, Any]:
+        operation_id = str(uuid.uuid4())
+        task_id = str(uuid.uuid4())
+        tool_call_id = f"sdk-{uuid.uuid4().hex[:12]}"
+        session_id = session.session_id
+
+        # --- Build user-friendly messages (matches non-SDK service) ---
+        if tool_name == "create_agent":
+            desc = args.get("description", "")
+            desc_preview = (desc[:100] + "...") if len(desc) > 100 else desc
+            pending_msg = (
+                f"Creating your agent: {desc_preview}"
+                if desc_preview
+                else "Creating agent... This may take a few minutes."
+            )
+            started_msg = (
+                "Agent creation started. You can close this tab - "
+                "check your library in a few minutes."
+            )
+        elif tool_name == "edit_agent":
+            changes = args.get("changes", "")
+            changes_preview = (changes[:100] + "...") if len(changes) > 100 else changes
+            pending_msg = (
+                f"Editing agent: {changes_preview}"
+                if changes_preview
+                else "Editing agent... This may take a few minutes."
+            )
+            started_msg = (
+                "Agent edit started. You can close this tab - "
+                "check your library in a few minutes."
+            )
+        else:
+            pending_msg = f"Running {tool_name}... This may take a few minutes."
+            started_msg = (
+                f"{tool_name} started. You can close this tab - "
+                "check back in a few minutes."
+            )
+
+        # --- Register task in Redis for SSE reconnection ---
+        await stream_registry.create_task(
+            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,
+        )
+
+        # --- Save OperationPendingResponse to chat history ---
+        pending_message = ChatMessage(
+            role="tool",
+            content=OperationPendingResponse(
+                message=pending_msg,
+                operation_id=operation_id,
+                tool_name=tool_name,
+            ).model_dump_json(),
+            tool_call_id=tool_call_id,
+        )
+        session.messages.append(pending_message)
+        await upsert_chat_session(session)
+
+        # --- Spawn background task (reuses non-SDK infrastructure) ---
+        bg_task = asyncio.create_task(
+            _execute_long_running_tool_with_streaming(
+                tool_name=tool_name,
+                parameters=args,
+                tool_call_id=tool_call_id,
+                operation_id=operation_id,
+                task_id=task_id,
+                session_id=session_id,
+                user_id=user_id,
+            )
+        )
+        _background_tasks.add(bg_task)
+        bg_task.add_done_callback(_background_tasks.discard)
+        await stream_registry.set_task_asyncio_task(task_id, bg_task)
+
+        logger.info(
+            f"[SDK] Long-running tool {tool_name} delegated to background "
+            f"(operation_id={operation_id}, task_id={task_id})"
+        )
+
+        # --- Return OperationStartedResponse as MCP tool result ---
+        # This flows through SDK → response adapter → frontend, triggering
+        # the loading widget with SSE reconnection support.
+        started_json = OperationStartedResponse(
+            message=started_msg,
+            operation_id=operation_id,
+            tool_name=tool_name,
+            task_id=task_id,
+        ).model_dump_json()
+
+        return {
+            "content": [{"type": "text", "text": started_json}],
+            "isError": False,
+        }
+
+    return _callback
+
+
+def _resolve_sdk_model() -> str | None:
+    """Resolve the model name for the Claude Agent SDK CLI.
+
+    Uses ``config.claude_agent_model`` if set, otherwise derives from
+    ``config.model`` by stripping the OpenRouter provider prefix (e.g.,
+    ``"anthropic/claude-opus-4.6"`` → ``"claude-opus-4.6"``).
+    """
+    if config.claude_agent_model:
+        return config.claude_agent_model
+    model = config.model
+    if "/" in model:
+        return model.split("/", 1)[1]
+    return model
+
+
+def _build_sdk_env() -> dict[str, str]:
+    """Build env vars for the SDK CLI process.
+
+    Routes API calls through OpenRouter (or a custom base_url) using
+    the same ``config.api_key`` / ``config.base_url`` as the non-SDK path.
+    This gives per-call token and cost tracking on the OpenRouter dashboard.
+
+    Only overrides ``ANTHROPIC_API_KEY`` when a valid proxy URL and auth
+    token are both present — otherwise returns an empty dict so the SDK
+    falls back to its default credentials.
+    """
+    env: dict[str, str] = {}
+    if config.api_key and config.base_url:
+        # Strip /v1 suffix — SDK expects the base URL without a version path
+        base = config.base_url.rstrip("/")
+        if base.endswith("/v1"):
+            base = base[:-3]
+        if not base or not base.startswith("http"):
+            # Invalid base_url — don't override SDK defaults
+            return env
+        env["ANTHROPIC_BASE_URL"] = base
+        env["ANTHROPIC_AUTH_TOKEN"] = config.api_key
+        # Must be explicitly empty so the CLI uses AUTH_TOKEN instead
+        env["ANTHROPIC_API_KEY"] = ""
+    return env
+
+
+def _make_sdk_cwd(session_id: str) -> str:
+    """Create a safe, session-specific working directory path.
+
+    Delegates to :func:`~backend.api.features.chat.tools.sandbox.make_session_path`
+    (single source of truth for path sanitization) and adds a defence-in-depth
+    assertion.
+    """
+    cwd = make_session_path(session_id)
+    # Defence-in-depth: normpath + startswith is a CodeQL-recognised sanitizer
+    cwd = os.path.normpath(cwd)
+    if not cwd.startswith(_SDK_CWD_PREFIX):
+        raise ValueError(f"SDK cwd escaped prefix: {cwd}")
+    return cwd
+
+
+def _cleanup_sdk_tool_results(cwd: str) -> None:
+    """Remove SDK tool-result files for a specific session working directory.
+
+    The SDK creates tool-result files under ~/.claude/projects/<encoded-cwd>/tool-results/.
+    We clean only the specific cwd's results to avoid race conditions between
+    concurrent sessions.
+
+    Security: cwd MUST be created by _make_sdk_cwd() which sanitizes session_id.
+    """
+    import shutil
+
+    # Security check 1: Validate cwd is under the expected prefix
+    normalized = os.path.normpath(cwd)
+    if not normalized.startswith(_SDK_CWD_PREFIX):
+        logger.warning(f"[SDK] Rejecting cleanup for invalid path: {cwd}")
+        return
+
+    # Security check 2: Ensure no path traversal in the normalized path
+    if ".." in normalized:
+        logger.warning(f"[SDK] Rejecting cleanup for traversal attempt: {cwd}")
+        return
+
+    # SDK encodes the cwd path by replacing '/' with '-'
+    encoded_cwd = normalized.replace("/", "-")
+
+    # Construct the project directory path (known-safe home expansion)
+    claude_projects = os.path.expanduser("~/.claude/projects")
+    project_dir = os.path.join(claude_projects, encoded_cwd)
+
+    # Security check 3: Validate project_dir is under ~/.claude/projects
+    project_dir = os.path.normpath(project_dir)
+    if not project_dir.startswith(claude_projects):
+        logger.warning(
+            f"[SDK] Rejecting cleanup for escaped project path: {project_dir}"
+        )
+        return
+
+    results_dir = os.path.join(project_dir, "tool-results")
+    if os.path.isdir(results_dir):
+        for filename in os.listdir(results_dir):
+            file_path = os.path.join(results_dir, filename)
+            try:
+                if os.path.isfile(file_path):
+                    os.remove(file_path)
+            except OSError:
+                pass
+
+    # Also clean up the temp cwd directory itself
+    try:
+        shutil.rmtree(normalized, ignore_errors=True)
+    except OSError:
+        pass
+
+
+async def _compress_conversation_history(
+    session: ChatSession,
+) -> list[ChatMessage]:
+    """Compress prior conversation messages if they exceed the token threshold.
+
+    Uses the shared compress_context() from prompt.py which supports:
+    - LLM summarization of old messages (keeps recent ones intact)
+    - Progressive content truncation as fallback
+    - Middle-out deletion as last resort
+
+    Returns the compressed prior messages (everything except the current message).
+    """
+    prior = session.messages[:-1]
+    if len(prior) < 2:
+        return prior
+
+    from backend.util.prompt import compress_context
+
+    # Convert ChatMessages to dicts for compress_context
+    messages_dict = []
+    for msg in prior:
+        msg_dict: dict[str, Any] = {"role": msg.role}
+        if msg.content:
+            msg_dict["content"] = msg.content
+        if msg.tool_calls:
+            msg_dict["tool_calls"] = msg.tool_calls
+        if msg.tool_call_id:
+            msg_dict["tool_call_id"] = msg.tool_call_id
+        messages_dict.append(msg_dict)
+
+    try:
+        import openai
+
+        async with openai.AsyncOpenAI(
+            api_key=config.api_key, base_url=config.base_url, timeout=30.0
+        ) as client:
+            result = await compress_context(
+                messages=messages_dict,
+                model=config.model,
+                client=client,
+            )
+    except Exception as e:
+        logger.warning(f"[SDK] Context compression with LLM failed: {e}")
+        # Fall back to truncation-only (no LLM summarization)
+        result = await compress_context(
+            messages=messages_dict,
+            model=config.model,
+            client=None,
+        )
+
+    if result.was_compacted:
+        logger.info(
+            f"[SDK] Context compacted: {result.original_token_count} -> "
+            f"{result.token_count} tokens "
+            f"({result.messages_summarized} summarized, "
+            f"{result.messages_dropped} dropped)"
+        )
+        # Convert compressed dicts back to ChatMessages
+        return [
+            ChatMessage(
+                role=m["role"],
+                content=m.get("content"),
+                tool_calls=m.get("tool_calls"),
+                tool_call_id=m.get("tool_call_id"),
+            )
+            for m in result.messages
+        ]
+
+    return prior
+
+
+def _format_conversation_context(messages: list[ChatMessage]) -> str | None:
+    """Format conversation messages into a context prefix for the user message.
+
+    Returns a string like:
+        <conversation_history>
+        User: hello
+        You responded: Hi! How can I help?
+        </conversation_history>
+
+    Returns None if there are no messages to format.
+    """
+    if not messages:
+        return None
+
+    lines: list[str] = []
+    for msg in messages:
+        if not msg.content:
+            continue
+        if msg.role == "user":
+            lines.append(f"User: {msg.content}")
+        elif msg.role == "assistant":
+            lines.append(f"You responded: {msg.content}")
+        # Skip tool messages — they're internal details
+
+    if not lines:
+        return None
+
+    return "<conversation_history>\n" + "\n".join(lines) + "\n</conversation_history>"
+
+
+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)
+                )
+                _background_tasks.add(task)
+                task.add_done_callback(_background_tasks.discard)
+
+    # Build system prompt (reuses non-SDK path with Langfuse support)
+    has_history = len(session.messages) > 1
+    system_prompt, _ = await _build_system_prompt(
+        user_id, has_conversation_history=has_history
+    )
+    system_prompt += _SDK_TOOL_SUPPLEMENT
+    message_id = str(uuid.uuid4())
+    task_id = str(uuid.uuid4())
+
+    yield StreamStart(messageId=message_id, taskId=task_id)
+
+    stream_completed = False
+    # Initialise sdk_cwd before the try so the finally can reference it
+    # even if _make_sdk_cwd raises (in that case it stays as "").
+    sdk_cwd = ""
+
+    try:
+        # Use a session-specific temp dir to avoid cleanup race conditions
+        # between concurrent sessions.
+        sdk_cwd = _make_sdk_cwd(session_id)
+        os.makedirs(sdk_cwd, exist_ok=True)
+
+        set_execution_context(
+            user_id,
+            session,
+            long_running_callback=_build_long_running_callback(user_id),
+        )
+        try:
+            from claude_agent_sdk import ClaudeAgentOptions, ClaudeSDKClient
+
+            # Fail fast when no API credentials are available at all
+            sdk_env = _build_sdk_env()
+            if not sdk_env and not os.environ.get("ANTHROPIC_API_KEY"):
+                raise RuntimeError(
+                    "No API key configured. Set OPEN_ROUTER_API_KEY "
+                    "(or CHAT_API_KEY) for OpenRouter routing, "
+                    "or ANTHROPIC_API_KEY for direct Anthropic access."
+                )
+
+            mcp_server = create_copilot_mcp_server()
+
+            sdk_model = _resolve_sdk_model()
+
+            security_hooks = create_security_hooks(
+                user_id,
+                sdk_cwd=sdk_cwd,
+                max_subtasks=config.claude_agent_max_subtasks,
+            )
+
+            options = ClaudeAgentOptions(
+                system_prompt=system_prompt,
+                mcp_servers={"copilot": mcp_server},  # type: ignore[arg-type]
+                allowed_tools=COPILOT_TOOL_NAMES,
+                hooks=security_hooks,  # type: ignore[arg-type]
+                cwd=sdk_cwd,
+                max_buffer_size=config.claude_agent_max_buffer_size,
+                # Only pass model/env when OpenRouter is configured
+                **({"model": sdk_model, "env": sdk_env} if sdk_env else {}),
+            )
+
+            adapter = SDKResponseAdapter(message_id=message_id)
+            adapter.set_task_id(task_id)
+
+            async with ClaudeSDKClient(options=options) as client:
+                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 ""
+
+                if not current_message.strip():
+                    yield StreamError(
+                        errorText="Message cannot be empty.",
+                        code="empty_prompt",
+                    )
+                    yield StreamFinish()
+                    return
+
+                # Build query with conversation history context.
+                # Compress history first to handle long conversations.
+                query_message = current_message
+                if len(session.messages) > 1:
+                    compressed = await _compress_conversation_history(session)
+                    history_context = _format_conversation_context(compressed)
+                    if history_context:
+                        query_message = (
+                            f"{history_context}\n\n"
+                            f"Now, the user says:\n{current_message}"
+                        )
+
+                logger.info(
+                    f"[SDK] Sending query: {current_message[:80]!r}"
+                    f" ({len(session.messages)} msgs in session)"
+                )
+                await client.query(query_message, session_id=session_id)
+
+                assistant_response = ChatMessage(role="assistant", content="")
+                accumulated_tool_calls: list[dict[str, Any]] = []
+                has_appended_assistant = False
+                has_tool_results = False
+
+                async for sdk_msg in client.receive_messages():
+                    logger.debug(
+                        f"[SDK] Received: {type(sdk_msg).__name__} "
+                        f"{getattr(sdk_msg, 'subtype', '')}"
+                    )
+                    for response in adapter.convert_message(sdk_msg):
+                        if isinstance(response, StreamStart):
+                            continue
+
+                        yield response
+
+                        if isinstance(response, StreamTextDelta):
+                            delta = response.delta or ""
+                            # After tool results, start a new assistant
+                            # message for the post-tool text.
+                            if has_tool_results and has_appended_assistant:
+                                assistant_response = ChatMessage(
+                                    role="assistant", content=delta
+                                )
+                                accumulated_tool_calls = []
+                                has_appended_assistant = False
+                                has_tool_results = False
+                                session.messages.append(assistant_response)
+                                has_appended_assistant = True
+                            else:
+                                assistant_response.content = (
+                                    assistant_response.content or ""
+                                ) + delta
+                                if not has_appended_assistant:
+                                    session.messages.append(assistant_response)
+                                    has_appended_assistant = True
+
+                        elif isinstance(response, StreamToolInputAvailable):
+                            accumulated_tool_calls.append(
+                                {
+                                    "id": response.toolCallId,
+                                    "type": "function",
+                                    "function": {
+                                        "name": response.toolName,
+                                        "arguments": json.dumps(response.input or {}),
+                                    },
+                                }
+                            )
+                            assistant_response.tool_calls = accumulated_tool_calls
+                            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
+
+                    if stream_completed:
+                        break
+
+                if (
+                    assistant_response.content or assistant_response.tool_calls
+                ) and not has_appended_assistant:
+                    session.messages.append(assistant_response)
+
+        except ImportError:
+            raise RuntimeError(
+                "claude-agent-sdk is not installed. "
+                "Disable SDK mode (CHAT_USE_CLAUDE_AGENT_SDK=false) "
+                "to use the OpenAI-compatible fallback."
+            )
+
+        await upsert_chat_session(session)
+        logger.debug(
+            f"[SDK] Session {session_id} saved with {len(session.messages)} messages"
+        )
+        if not stream_completed:
+            yield StreamFinish()
+
+    except Exception as e:
+        logger.error(f"[SDK] Error: {e}", exc_info=True)
+        try:
+            await upsert_chat_session(session)
+        except Exception as save_err:
+            logger.error(f"[SDK] Failed to save session on error: {save_err}")
+        yield StreamError(
+            errorText="An error occurred. Please try again.",
+            code="sdk_error",
+        )
+        yield StreamFinish()
+    finally:
+        if sdk_cwd:
+            _cleanup_sdk_tool_results(sdk_cwd)
+
+
+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 +1,320 @@
+"""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.
+
+Long-running tools (``is_long_running=True``) are delegated to the non-SDK
+background infrastructure (stream_registry, Redis persistence, SSE reconnection)
+via a callback provided by the service layer.  This avoids wasteful SDK polling
+and makes results survive page refreshes.
+"""
+
+import json
+import logging
+import os
+import uuid
+from collections.abc import Awaitable, Callable
+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__)
+
+# Allowed base directory for the Read tool (SDK saves oversized tool results here).
+# Restricted to ~/.claude/projects/ and further validated to require "tool-results"
+# in the path — prevents reading settings, credentials, or other sensitive files.
+_SDK_PROJECTS_DIR = os.path.expanduser("~/.claude/projects/")
+
+# MCP server naming - the SDK prefixes tool names as "mcp__{server_name}__{tool}"
+MCP_SERVER_NAME = "copilot"
+MCP_TOOL_PREFIX = f"mcp__{MCP_SERVER_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
+)
+# Stash for MCP tool outputs before the SDK potentially truncates them.
+# Keyed by tool_name → full output string. Consumed (popped) by the
+# response adapter when it builds StreamToolOutputAvailable.
+_pending_tool_outputs: ContextVar[dict[str, str]] = ContextVar(
+    "pending_tool_outputs", default=None  # type: ignore[arg-type]
+)
+
+# Callback type for delegating long-running tools to the non-SDK infrastructure.
+# Args: (tool_name, arguments, session) → MCP-formatted response dict.
+LongRunningCallback = Callable[
+    [str, dict[str, Any], ChatSession], Awaitable[dict[str, Any]]
+]
+
+# ContextVar so the service layer can inject the callback per-request.
+_long_running_callback: ContextVar[LongRunningCallback | None] = ContextVar(
+    "long_running_callback", default=None
+)
+
+
+def set_execution_context(
+    user_id: str | None,
+    session: ChatSession,
+    long_running_callback: LongRunningCallback | 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.
+
+    Args:
+        user_id: Current user's ID.
+        session: Current chat session.
+        long_running_callback: Optional callback to delegate long-running tools
+            to the non-SDK background infrastructure (stream_registry + Redis).
+    """
+    _current_user_id.set(user_id)
+    _current_session.set(session)
+    _pending_tool_outputs.set({})
+    _long_running_callback.set(long_running_callback)
+
+
+def get_execution_context() -> tuple[str | None, ChatSession | None]:
+    """Get the current execution context."""
+    return (
+        _current_user_id.get(),
+        _current_session.get(),
+    )
+
+
+def pop_pending_tool_output(tool_name: str) -> str | None:
+    """Pop and return the stashed full output for *tool_name*.
+
+    The SDK CLI may truncate large tool results (writing them to disk and
+    replacing the content with a file reference). This stash keeps the
+    original MCP output so the response adapter can forward it to the
+    frontend for proper widget rendering.
+
+    Returns ``None`` if nothing was stashed for *tool_name*.
+    """
+    pending = _pending_tool_outputs.get(None)
+    if pending is None:
+        return None
+    return pending.pop(tool_name, None)
+
+
+async def _execute_tool_sync(
+    base_tool: BaseTool,
+    user_id: str | None,
+    session: ChatSession,
+    args: dict[str, Any],
+) -> dict[str, Any]:
+    """Execute a tool synchronously and return MCP-formatted response."""
+    effective_id = f"sdk-{uuid.uuid4().hex[:12]}"
+    result = await base_tool.execute(
+        user_id=user_id,
+        session=session,
+        tool_call_id=effective_id,
+        **args,
+    )
+
+    text = (
+        result.output if isinstance(result.output, str) else json.dumps(result.output)
+    )
+
+    # Stash the full output before the SDK potentially truncates it.
+    pending = _pending_tool_outputs.get(None)
+    if pending is not None:
+        pending[base_tool.name] = text
+
+    return {
+        "content": [{"type": "text", "text": text}],
+        "isError": not result.success,
+    }
+
+
+def _mcp_error(message: str) -> dict[str, Any]:
+    return {
+        "content": [
+            {"type": "text", "text": json.dumps({"error": message, "type": "error"})}
+        ],
+        "isError": True,
+    }
+
+
+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.
+
+    Long-running tools (``is_long_running=True``) are delegated to the
+    non-SDK background infrastructure via a callback set in the execution
+    context.  The callback persists the operation in Redis (stream_registry)
+    so results survive page refreshes and pod restarts.
+    """
+
+    async def tool_handler(args: dict[str, Any]) -> dict[str, Any]:
+        """Execute the wrapped tool and return MCP-formatted response."""
+        user_id, session = get_execution_context()
+
+        if session is None:
+            return _mcp_error("No session context available")
+
+        # --- Long-running: delegate to non-SDK background infrastructure ---
+        if base_tool.is_long_running:
+            callback = _long_running_callback.get(None)
+            if callback:
+                try:
+                    return await callback(base_tool.name, args, session)
+                except Exception as e:
+                    logger.error(
+                        f"Long-running callback failed for {base_tool.name}: {e}",
+                        exc_info=True,
+                    )
+                    return _mcp_error(f"Failed to start {base_tool.name}: {e}")
+            # No callback — fall through to synchronous execution
+            logger.warning(
+                f"[SDK] No long-running callback for {base_tool.name}, "
+                f"executing synchronously (may block)"
+            )
+
+        # --- Normal (fast) tool: execute synchronously ---
+        try:
+            return await _execute_tool_sync(base_tool, user_id, session, args)
+        except Exception as e:
+            logger.error(f"Error executing tool {base_tool.name}: {e}", exc_info=True)
+            return _mcp_error(f"Failed to execute {base_tool.name}: {e}")
+
+    return tool_handler
+
+
+def _build_input_schema(base_tool: BaseTool) -> dict[str, Any]:
+    """Build a JSON Schema input schema for a tool."""
+    return {
+        "type": "object",
+        "properties": base_tool.parameters.get("properties", {}),
+        "required": base_tool.parameters.get("required", []),
+    }
+
+
+async def _read_file_handler(args: dict[str, Any]) -> dict[str, Any]:
+    """Read a file with optional offset/limit. Restricted to SDK working directory.
+
+    After reading, the file is deleted to prevent accumulation in long-running pods.
+    """
+    file_path = args.get("file_path", "")
+    offset = args.get("offset", 0)
+    limit = args.get("limit", 2000)
+
+    # Security: only allow reads under ~/.claude/projects/**/tool-results/
+    real_path = os.path.realpath(file_path)
+    if not real_path.startswith(_SDK_PROJECTS_DIR) or "tool-results" not in real_path:
+        return {
+            "content": [{"type": "text", "text": f"Access denied: {file_path}"}],
+            "isError": True,
+        }
+
+    try:
+        with open(real_path) as f:
+            lines = f.readlines()
+        selected = lines[offset : offset + limit]
+        content = "".join(selected)
+        return {"content": [{"type": "text", "text": content}], "isError": False}
+    except FileNotFoundError:
+        return {
+            "content": [{"type": "text", "text": f"File not found: {file_path}"}],
+            "isError": True,
+        }
+    except Exception as e:
+        return {
+            "content": [{"type": "text", "text": f"Error reading file: {e}"}],
+            "isError": True,
+        }
+
+
+_READ_TOOL_NAME = "Read"
+_READ_TOOL_DESCRIPTION = (
+    "Read a file from the local filesystem. "
+    "Use offset and limit to read specific line ranges for large files."
+)
+_READ_TOOL_SCHEMA = {
+    "type": "object",
+    "properties": {
+        "file_path": {
+            "type": "string",
+            "description": "The absolute path to the file to read",
+        },
+        "offset": {
+            "type": "integer",
+            "description": "Line number to start reading from (0-indexed). Default: 0",
+        },
+        "limit": {
+            "type": "integer",
+            "description": "Number of lines to read. Default: 2000",
+        },
+    },
+    "required": ["file_path"],
+}
+
+
+# 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():
+            handler = create_tool_handler(base_tool)
+            decorated = tool(
+                tool_name,
+                base_tool.description,
+                _build_input_schema(base_tool),
+            )(handler)
+            sdk_tools.append(decorated)
+
+        # Add the Read tool so the SDK can read back oversized tool results
+        read_tool = tool(
+            _READ_TOOL_NAME,
+            _READ_TOOL_DESCRIPTION,
+            _READ_TOOL_SCHEMA,
+        )(_read_file_handler)
+        sdk_tools.append(read_tool)
+
+        server = create_sdk_mcp_server(
+            name=MCP_SERVER_NAME,
+            version="1.0.0",
+            tools=sdk_tools,
+        )
+
+        return server
+
+    except ImportError:
+        # Let ImportError propagate so service.py handles the fallback
+        raise
+
+
+# SDK built-in tools allowed within the workspace directory.
+# Security hooks validate that file paths stay within sdk_cwd.
+# Bash is NOT included — use the sandboxed MCP bash_exec tool instead,
+# which provides kernel-level network isolation via unshare --net.
+# Task allows spawning sub-agents (rate-limited by security hooks).
+_SDK_BUILTIN_TOOLS = ["Read", "Write", "Edit", "Glob", "Grep", "Task"]
+
+# List of tool names for allowed_tools configuration
+# Include MCP tools, the MCP Read tool for oversized results,
+# and SDK built-in file tools for workspace operations.
+COPILOT_TOOL_NAMES = [
+    *[f"{MCP_TOOL_PREFIX}{name}" for name in TOOL_REGISTRY.keys()],
+    f"{MCP_TOOL_PREFIX}{_READ_TOOL_NAME}",
+    *_SDK_BUILTIN_TOOLS,
+]

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

@@ -245,12 +245,16 @@ async def _get_system_prompt_template(context: str) -> str:
     return DEFAULT_SYSTEM_PROMPT.format(users_information=context)
 
 
-async def _build_system_prompt(user_id: str | None) -> tuple[str, Any]:
+async def _build_system_prompt(
+    user_id: str | None, has_conversation_history: bool = False
+) -> tuple[str, Any]:
     """Build the full system prompt including business understanding if available.
 
     Args:
-        user_id: The user ID for fetching business understanding
+        user_id: The user ID for fetching business understanding.
-                     If "default" and this is the user's first session, will use "onboarding" instead.
+        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).
 
     Returns:
         Tuple of (compiled prompt string, business understanding object)
@@ -266,6 +270,8 @@ async def _build_system_prompt(user_id: str | None) -> tuple[str, Any]:
 
     if understanding:
         context = format_understanding_for_prompt(understanding)
+    elif has_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"
 
@@ -374,7 +380,6 @@ async def stream_chat_completion(
 
     Raises:
         NotFoundError: If session_id is invalid
-        ValueError: If max_context_messages is exceeded
 
     """
     completion_start = time.monotonic()
@@ -459,8 +464,9 @@ async def stream_chat_completion(
 
     # Generate title for new sessions on first user message (non-blocking)
     # Check: is_user_message, no title yet, and this is the first user message
-    if is_user_message and message and not session.title:
+    user_messages = [m for m in session.messages if m.role == "user"]
-        user_messages = [m for m in session.messages if m.role == "user"]
+    first_user_msg = message or (user_messages[0].content if user_messages else None)
+    if is_user_message and first_user_msg and not session.title:
         if len(user_messages) == 1:
             # First user message - generate title in background
             import asyncio
@@ -468,7 +474,7 @@ async def stream_chat_completion(
             # Capture only the values we need (not the session object) to avoid
             # stale data issues when the main flow modifies the session
             captured_session_id = session_id
-            captured_message = message
+            captured_message = first_user_msg
             captured_user_id = user_id
 
             async def _update_title():
@@ -1237,7 +1243,7 @@ async def _stream_chat_chunks(
 
                 total_time = (time_module.perf_counter() - stream_chunks_start) * 1000
                 logger.info(
-                    f"[TIMING] _stream_chat_chunks COMPLETED in {total_time/1000:.1f}s; "
+                    f"[TIMING] _stream_chat_chunks COMPLETED in {total_time / 1000:.1f}s; "
                     f"session={session.session_id}, user={session.user_id}",
                     extra={"json_fields": {**log_meta, "total_time_ms": total_time}},
                 )

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

@@ -814,6 +814,28 @@ async def get_active_task_for_session(
                 if task_user_id and user_id != task_user_id:
                     continue
 
+                # Auto-expire stale tasks that exceeded stream_timeout
+                created_at_str = meta.get("created_at", "")
+                if created_at_str:
+                    try:
+                        created_at = datetime.fromisoformat(created_at_str)
+                        age_seconds = (
+                            datetime.now(timezone.utc) - created_at
+                        ).total_seconds()
+                        if age_seconds > config.stream_timeout:
+                            logger.warning(
+                                f"[TASK_LOOKUP] Auto-expiring stale task {task_id[:8]}... "
+                                f"(age={age_seconds:.0f}s > timeout={config.stream_timeout}s)"
+                            )
+                            await mark_task_completed(task_id, "failed")
+                            continue
+                    except (ValueError, TypeError):
+                        pass
+
+                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"

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

@@ -9,6 +9,8 @@ from backend.api.features.chat.tracking import track_tool_called
 from .add_understanding import AddUnderstandingTool
 from .agent_output import AgentOutputTool
 from .base import BaseTool
+from .bash_exec import BashExecTool
+from .check_operation_status import CheckOperationStatusTool
 from .create_agent import CreateAgentTool
 from .customize_agent import CustomizeAgentTool
 from .edit_agent import EditAgentTool
@@ -19,6 +21,7 @@ from .get_doc_page import GetDocPageTool
 from .run_agent import RunAgentTool
 from .run_block import RunBlockTool
 from .search_docs import SearchDocsTool
+from .web_fetch import WebFetchTool
 from .workspace_files import (
     DeleteWorkspaceFileTool,
     ListWorkspaceFilesTool,
@@ -43,9 +46,14 @@ TOOL_REGISTRY: dict[str, BaseTool] = {
     "run_agent": RunAgentTool(),
     "run_block": RunBlockTool(),
     "view_agent_output": AgentOutputTool(),
+    "check_operation_status": CheckOperationStatusTool(),
     "search_docs": SearchDocsTool(),
     "get_doc_page": GetDocPageTool(),
-    # Workspace tools for CoPilot file operations
+    # Web fetch for safe URL retrieval
+    "web_fetch": WebFetchTool(),
+    # Sandboxed code execution (bubblewrap)
+    "bash_exec": BashExecTool(),
+    # Persistent workspace tools (cloud storage, survives across sessions)
     "list_workspace_files": ListWorkspaceFilesTool(),
     "read_workspace_file": ReadWorkspaceFileTool(),
     "write_workspace_file": WriteWorkspaceFileTool(),

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

@@ -0,0 +1,131 @@
+"""Bash execution tool — run shell commands in a bubblewrap sandbox.
+
+Full Bash scripting is allowed (loops, conditionals, pipes, functions, etc.).
+Safety comes from OS-level isolation (bubblewrap): only system dirs visible
+read-only, writable workspace only, clean env, no network.
+
+Requires bubblewrap (``bwrap``) — the tool is disabled when bwrap is not
+available (e.g. macOS development).
+"""
+
+import logging
+from typing import Any
+
+from backend.api.features.chat.model import ChatSession
+from backend.api.features.chat.tools.base import BaseTool
+from backend.api.features.chat.tools.models import (
+    BashExecResponse,
+    ErrorResponse,
+    ToolResponseBase,
+)
+from backend.api.features.chat.tools.sandbox import (
+    get_workspace_dir,
+    has_full_sandbox,
+    run_sandboxed,
+)
+
+logger = logging.getLogger(__name__)
+
+
+class BashExecTool(BaseTool):
+    """Execute Bash commands in a bubblewrap sandbox."""
+
+    @property
+    def name(self) -> str:
+        return "bash_exec"
+
+    @property
+    def description(self) -> str:
+        if not has_full_sandbox():
+            return (
+                "Bash execution is DISABLED — bubblewrap sandbox is not "
+                "available on this platform. Do not call this tool."
+            )
+        return (
+            "Execute a Bash command or script in a bubblewrap sandbox. "
+            "Full Bash scripting is supported (loops, conditionals, pipes, "
+            "functions, etc.). "
+            "The sandbox shares the same working directory as the SDK Read/Write "
+            "tools — files created by either are accessible to both. "
+            "SECURITY: Only system directories (/usr, /bin, /lib, /etc) are "
+            "visible read-only, the per-session workspace is the only writable "
+            "path, environment variables are wiped (no secrets), all network "
+            "access is blocked at the kernel level, and resource limits are "
+            "enforced (max 64 processes, 512MB memory, 50MB file size). "
+            "Application code, configs, and other directories are NOT accessible. "
+            "To fetch web content, use the web_fetch tool instead. "
+            "Execution is killed after the timeout (default 30s, max 120s). "
+            "Returns stdout and stderr. "
+            "Useful for file manipulation, data processing with Unix tools "
+            "(grep, awk, sed, jq, etc.), and running shell scripts."
+        )
+
+    @property
+    def parameters(self) -> dict[str, Any]:
+        return {
+            "type": "object",
+            "properties": {
+                "command": {
+                    "type": "string",
+                    "description": "Bash command or script to execute.",
+                },
+                "timeout": {
+                    "type": "integer",
+                    "description": (
+                        "Max execution time in seconds (default 30, max 120)."
+                    ),
+                    "default": 30,
+                },
+            },
+            "required": ["command"],
+        }
+
+    @property
+    def requires_auth(self) -> bool:
+        return False
+
+    async def _execute(
+        self,
+        user_id: str | None,
+        session: ChatSession,
+        **kwargs: Any,
+    ) -> ToolResponseBase:
+        session_id = session.session_id if session else None
+
+        if not has_full_sandbox():
+            return ErrorResponse(
+                message="bash_exec requires bubblewrap sandbox (Linux only).",
+                error="sandbox_unavailable",
+                session_id=session_id,
+            )
+
+        command: str = (kwargs.get("command") or "").strip()
+        timeout: int = kwargs.get("timeout", 30)
+
+        if not command:
+            return ErrorResponse(
+                message="No command provided.",
+                error="empty_command",
+                session_id=session_id,
+            )
+
+        workspace = get_workspace_dir(session_id or "default")
+
+        stdout, stderr, exit_code, timed_out = await run_sandboxed(
+            command=["bash", "-c", command],
+            cwd=workspace,
+            timeout=timeout,
+        )
+
+        return BashExecResponse(
+            message=(
+                "Execution timed out"
+                if timed_out
+                else f"Command executed (exit {exit_code})"
+            ),
+            stdout=stdout,
+            stderr=stderr,
+            exit_code=exit_code,
+            timed_out=timed_out,
+            session_id=session_id,
+        )

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

@@ -0,0 +1,127 @@
+"""CheckOperationStatusTool — query the status of a long-running operation."""
+
+import logging
+from typing import Any
+
+from backend.api.features.chat.model import ChatSession
+from backend.api.features.chat.tools.base import BaseTool
+from backend.api.features.chat.tools.models import (
+    ErrorResponse,
+    ResponseType,
+    ToolResponseBase,
+)
+
+logger = logging.getLogger(__name__)
+
+
+class OperationStatusResponse(ToolResponseBase):
+    """Response for check_operation_status tool."""
+
+    type: ResponseType = ResponseType.OPERATION_STATUS
+    task_id: str
+    operation_id: str
+    status: str  # "running", "completed", "failed"
+    tool_name: str | None = None
+    message: str = ""
+
+
+class CheckOperationStatusTool(BaseTool):
+    """Check the status of a long-running operation (create_agent, edit_agent, etc.).
+
+    The CoPilot uses this tool to report back to the user whether an
+    operation that was started earlier has completed, failed, or is still
+    running.
+    """
+
+    @property
+    def name(self) -> str:
+        return "check_operation_status"
+
+    @property
+    def description(self) -> str:
+        return (
+            "Check the current status of a long-running operation such as "
+            "create_agent or edit_agent. Accepts either an operation_id or "
+            "task_id from a previous operation_started response. "
+            "Returns the current status: running, completed, or failed."
+        )
+
+    @property
+    def parameters(self) -> dict[str, Any]:
+        return {
+            "type": "object",
+            "properties": {
+                "operation_id": {
+                    "type": "string",
+                    "description": (
+                        "The operation_id from an operation_started response."
+                    ),
+                },
+                "task_id": {
+                    "type": "string",
+                    "description": (
+                        "The task_id from an operation_started response. "
+                        "Used as fallback if operation_id is not provided."
+                    ),
+                },
+            },
+            "required": [],
+        }
+
+    @property
+    def requires_auth(self) -> bool:
+        return False
+
+    async def _execute(
+        self,
+        user_id: str | None,
+        session: ChatSession,
+        **kwargs,
+    ) -> ToolResponseBase:
+        from backend.api.features.chat import stream_registry
+
+        operation_id: str = kwargs.get("operation_id", "").strip()
+        task_id: str = kwargs.get("task_id", "").strip()
+
+        if not operation_id and not task_id:
+            return ErrorResponse(
+                message="Please provide an operation_id or task_id.",
+                error="missing_parameter",
+            )
+
+        task = None
+        if operation_id:
+            task = await stream_registry.find_task_by_operation_id(operation_id)
+        if task is None and task_id:
+            task = await stream_registry.get_task(task_id)
+
+        if task is None:
+            # Task not in Redis — it may have already expired (TTL).
+            # Check conversation history for the result instead.
+            return ErrorResponse(
+                message=(
+                    "Operation not found — it may have already completed and "
+                    "expired from the status tracker. Check the conversation "
+                    "history for the result."
+                ),
+                error="not_found",
+            )
+
+        status_messages = {
+            "running": (
+                f"The {task.tool_name or 'operation'} is still running. "
+                "Please wait for it to complete."
+            ),
+            "completed": (
+                f"The {task.tool_name or 'operation'} has completed successfully."
+            ),
+            "failed": f"The {task.tool_name or 'operation'} has failed.",
+        }
+
+        return OperationStatusResponse(
+            task_id=task.task_id,
+            operation_id=task.operation_id,
+            status=task.status,
+            tool_name=task.tool_name,
+            message=status_messages.get(task.status, f"Status: {task.status}"),
+        )

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

@@ -40,6 +40,12 @@ class ResponseType(str, Enum):
     OPERATION_IN_PROGRESS = "operation_in_progress"
     # Input validation
     INPUT_VALIDATION_ERROR = "input_validation_error"
+    # Web fetch
+    WEB_FETCH = "web_fetch"
+    # Code execution
+    BASH_EXEC = "bash_exec"
+    # Operation status check
+    OPERATION_STATUS = "operation_status"
 
 
 # Base response model
@@ -335,11 +341,17 @@ class BlockInfoSummary(BaseModel):
     name: str
     description: str
     categories: list[str]
-    input_schema: dict[str, Any]
+    input_schema: dict[str, Any] = Field(
-    output_schema: dict[str, Any]
+        default_factory=dict,
+        description="Full JSON schema for block inputs",
+    )
+    output_schema: dict[str, Any] = Field(
+        default_factory=dict,
+        description="Full JSON schema for block outputs",
+    )
     required_inputs: list[BlockInputFieldInfo] = Field(
         default_factory=list,
-        description="List of required input fields for this block",
+        description="List of input fields for this block",
     )
 
 
@@ -352,7 +364,7 @@ class BlockListResponse(ToolResponseBase):
     query: str
     usage_hint: str = Field(
         default="To execute a block, call run_block with block_id set to the block's "
-        "'id' field and input_data containing the required fields from input_schema."
+        "'id' field and input_data containing the fields listed in required_inputs."
     )
 
 
@@ -421,3 +433,24 @@ class AsyncProcessingResponse(ToolResponseBase):
     status: str = "accepted"  # Must be "accepted" for detection
     operation_id: str | None = None
     task_id: str | None = None
+
+
+class WebFetchResponse(ToolResponseBase):
+    """Response for web_fetch tool."""
+
+    type: ResponseType = ResponseType.WEB_FETCH
+    url: str
+    status_code: int
+    content_type: str
+    content: str
+    truncated: bool = False
+
+
+class BashExecResponse(ToolResponseBase):
+    """Response for bash_exec tool."""
+
+    type: ResponseType = ResponseType.BASH_EXEC
+    stdout: str
+    stderr: str
+    exit_code: int
+    timed_out: bool = False

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

@@ -0,0 +1,267 @@
+"""Sandbox execution utilities for code execution tools.
+
+Provides filesystem + network isolated command execution using **bubblewrap**
+(``bwrap``): whitelist-only filesystem (only system dirs visible read-only),
+writable workspace only, clean environment, network blocked.
+
+Tools that call :func:`run_sandboxed` must first check :func:`has_full_sandbox`
+and refuse to run if bubblewrap is not available.
+"""
+
+import asyncio
+import logging
+import os
+import platform
+import shutil
+
+logger = logging.getLogger(__name__)
+
+# Output limits — prevent blowing up LLM context
+_MAX_OUTPUT_CHARS = 50_000
+_DEFAULT_TIMEOUT = 30
+_MAX_TIMEOUT = 120
+
+
+# ---------------------------------------------------------------------------
+# Sandbox capability detection (cached at first call)
+# ---------------------------------------------------------------------------
+
+_BWRAP_AVAILABLE: bool | None = None
+
+
+def has_full_sandbox() -> bool:
+    """Return True if bubblewrap is available (filesystem + network isolation).
+
+    On non-Linux platforms (macOS), always returns False.
+    """
+    global _BWRAP_AVAILABLE
+    if _BWRAP_AVAILABLE is None:
+        _BWRAP_AVAILABLE = (
+            platform.system() == "Linux" and shutil.which("bwrap") is not None
+        )
+    return _BWRAP_AVAILABLE
+
+
+WORKSPACE_PREFIX = "/tmp/copilot-"
+
+
+def make_session_path(session_id: str) -> str:
+    """Build a sanitized, session-specific path under :data:`WORKSPACE_PREFIX`.
+
+    Shared by both the SDK working-directory setup and the sandbox tools so
+    they always resolve to the same directory for a given session.
+
+    Steps:
+        1. Strip all characters except ``[A-Za-z0-9-]``.
+        2. Construct ``/tmp/copilot-<safe_id>``.
+        3. Validate via ``os.path.normpath`` + ``startswith`` (CodeQL-recognised
+           sanitizer) to prevent path traversal.
+
+    Raises:
+        ValueError: If the resulting path escapes the prefix.
+    """
+    import re
+
+    safe_id = re.sub(r"[^A-Za-z0-9-]", "", session_id)
+    if not safe_id:
+        safe_id = "default"
+    path = os.path.normpath(f"{WORKSPACE_PREFIX}{safe_id}")
+    if not path.startswith(WORKSPACE_PREFIX):
+        raise ValueError(f"Session path escaped prefix: {path}")
+    return path
+
+
+def get_workspace_dir(session_id: str) -> str:
+    """Get or create the workspace directory for a session.
+
+    Uses :func:`make_session_path` — the same path the SDK uses — so that
+    bash_exec shares the workspace with the SDK file tools.
+    """
+    workspace = make_session_path(session_id)
+    os.makedirs(workspace, exist_ok=True)
+    return workspace
+
+
+# ---------------------------------------------------------------------------
+# Bubblewrap command builder
+# ---------------------------------------------------------------------------
+
+# System directories mounted read-only inside the sandbox.
+# ONLY these are visible — /app, /root, /home, /opt, /var etc. are NOT accessible.
+_SYSTEM_RO_BINDS = [
+    "/usr",  # binaries, libraries, Python interpreter
+    "/etc",  # system config: ld.so, locale, passwd, alternatives
+]
+
+# Compat paths: symlinks to /usr/* on modern Debian, real dirs on older systems.
+# On Debian 13 these are symlinks (e.g. /bin -> usr/bin).  bwrap --ro-bind
+# can't create a symlink target, so we detect and use --symlink instead.
+# /lib64 is critical: the ELF dynamic linker lives at /lib64/ld-linux-x86-64.so.2.
+_COMPAT_PATHS = [
+    ("/bin", "usr/bin"),  # -> /usr/bin on Debian 13
+    ("/sbin", "usr/sbin"),  # -> /usr/sbin on Debian 13
+    ("/lib", "usr/lib"),  # -> /usr/lib on Debian 13
+    ("/lib64", "usr/lib64"),  # 64-bit libraries / ELF interpreter
+]
+
+# Resource limits to prevent fork bombs, memory exhaustion, and disk abuse.
+# Applied via ulimit inside the sandbox before exec'ing the user command.
+_RESOURCE_LIMITS = (
+    "ulimit -u 64"  # max 64 processes  (prevents fork bombs)
+    " -v 524288"  # 512 MB virtual memory
+    " -f 51200"  # 50 MB max file size  (1024-byte blocks)
+    " -n 256"  # 256 open file descriptors
+    " 2>/dev/null"
+)
+
+
+def _build_bwrap_command(
+    command: list[str], cwd: str, env: dict[str, str]
+) -> list[str]:
+    """Build a bubblewrap command with strict filesystem + network isolation.
+
+    Security model:
+    - **Whitelist-only filesystem**: only system directories (``/usr``, ``/etc``,
+      ``/bin``, ``/lib``) are mounted read-only.  Application code (``/app``),
+      home directories, ``/var``, ``/opt``, etc. are NOT accessible at all.
+    - **Writable workspace only**: the per-session workspace is the sole
+      writable path.
+    - **Clean environment**: ``--clearenv`` wipes all inherited env vars.
+      Only the explicitly-passed safe env vars are set inside the sandbox.
+    - **Network isolation**: ``--unshare-net`` blocks all network access.
+    - **Resource limits**: ulimit caps on processes (64), memory (512MB),
+      file size (50MB), and open FDs (256) to prevent fork bombs and abuse.
+    - **New session**: prevents terminal control escape.
+    - **Die with parent**: prevents orphaned sandbox processes.
+    """
+    cmd = [
+        "bwrap",
+        # Create a new user namespace so bwrap can set up sandboxing
+        # inside unprivileged Docker containers (no CAP_SYS_ADMIN needed).
+        "--unshare-user",
+        # Wipe all inherited environment variables (API keys, secrets, etc.)
+        "--clearenv",
+    ]
+
+    # Set only the safe env vars inside the sandbox
+    for key, value in env.items():
+        cmd.extend(["--setenv", key, value])
+
+    # System directories: read-only
+    for path in _SYSTEM_RO_BINDS:
+        cmd.extend(["--ro-bind", path, path])
+
+    # Compat paths: use --symlink when host path is a symlink (Debian 13),
+    # --ro-bind when it's a real directory (older distros).
+    for path, symlink_target in _COMPAT_PATHS:
+        if os.path.islink(path):
+            cmd.extend(["--symlink", symlink_target, path])
+        elif os.path.exists(path):
+            cmd.extend(["--ro-bind", path, path])
+
+    # Wrap the user command with resource limits:
+    #   sh -c 'ulimit ...; exec "$@"' -- <original command>
+    # `exec "$@"` replaces the shell so there's no extra process overhead,
+    # and properly handles arguments with spaces.
+    limited_command = [
+        "sh",
+        "-c",
+        f'{_RESOURCE_LIMITS}; exec "$@"',
+        "--",
+        *command,
+    ]
+
+    cmd.extend(
+        [
+            # Fresh virtual filesystems
+            "--dev",
+            "/dev",
+            "--proc",
+            "/proc",
+            "--tmpfs",
+            "/tmp",
+            # Workspace bind AFTER --tmpfs /tmp so it's visible through the tmpfs.
+            # (workspace lives under /tmp/copilot-<session>)
+            "--bind",
+            cwd,
+            cwd,
+            # Isolation
+            "--unshare-net",
+            "--die-with-parent",
+            "--new-session",
+            "--chdir",
+            cwd,
+            "--",
+            *limited_command,
+        ]
+    )
+
+    return cmd
+
+
+# ---------------------------------------------------------------------------
+# Public API
+# ---------------------------------------------------------------------------
+
+
+async def run_sandboxed(
+    command: list[str],
+    cwd: str,
+    timeout: int = _DEFAULT_TIMEOUT,
+    env: dict[str, str] | None = None,
+) -> tuple[str, str, int, bool]:
+    """Run a command inside a bubblewrap sandbox.
+
+    Callers **must** check :func:`has_full_sandbox` before calling this
+    function.  If bubblewrap is not available, this function raises
+    :class:`RuntimeError` rather than running unsandboxed.
+
+    Returns:
+        (stdout, stderr, exit_code, timed_out)
+    """
+    if not has_full_sandbox():
+        raise RuntimeError(
+            "run_sandboxed() requires bubblewrap but bwrap is not available. "
+            "Callers must check has_full_sandbox() before calling this function."
+        )
+
+    timeout = min(max(timeout, 1), _MAX_TIMEOUT)
+
+    safe_env = {
+        "PATH": "/usr/local/bin:/usr/bin:/bin",
+        "HOME": cwd,
+        "TMPDIR": cwd,
+        "LANG": "en_US.UTF-8",
+        "PYTHONDONTWRITEBYTECODE": "1",
+        "PYTHONIOENCODING": "utf-8",
+    }
+    if env:
+        safe_env.update(env)
+
+    full_command = _build_bwrap_command(command, cwd, safe_env)
+
+    try:
+        proc = await asyncio.create_subprocess_exec(
+            *full_command,
+            stdout=asyncio.subprocess.PIPE,
+            stderr=asyncio.subprocess.PIPE,
+            cwd=cwd,
+            env=safe_env,
+        )
+
+        try:
+            stdout_bytes, stderr_bytes = await asyncio.wait_for(
+                proc.communicate(), timeout=timeout
+            )
+            stdout = stdout_bytes.decode("utf-8", errors="replace")[:_MAX_OUTPUT_CHARS]
+            stderr = stderr_bytes.decode("utf-8", errors="replace")[:_MAX_OUTPUT_CHARS]
+            return stdout, stderr, proc.returncode or 0, False
+        except asyncio.TimeoutError:
+            proc.kill()
+            await proc.communicate()
+            return "", f"Execution timed out after {timeout}s", -1, True
+
+    except RuntimeError:
+        raise
+    except Exception as e:
+        return "", f"Sandbox error: {e}", -1, False

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

@@ -0,0 +1,156 @@
+"""Web fetch tool — safely retrieve public web page content."""
+
+import logging
+from typing import Any
+
+import aiohttp
+import html2text
+
+from backend.api.features.chat.model import ChatSession
+from backend.api.features.chat.tools.base import BaseTool
+from backend.api.features.chat.tools.models import (
+    ErrorResponse,
+    ToolResponseBase,
+    WebFetchResponse,
+)
+from backend.util.request import Requests
+
+logger = logging.getLogger(__name__)
+
+# Limits
+_MAX_CONTENT_BYTES = 102_400  # 100 KB download cap
+_MAX_OUTPUT_CHARS = 50_000  # 50K char truncation for LLM context
+_REQUEST_TIMEOUT = aiohttp.ClientTimeout(total=15)
+
+# Content types we'll read as text
+_TEXT_CONTENT_TYPES = {
+    "text/html",
+    "text/plain",
+    "text/xml",
+    "text/csv",
+    "text/markdown",
+    "application/json",
+    "application/xml",
+    "application/xhtml+xml",
+    "application/rss+xml",
+    "application/atom+xml",
+}
+
+
+def _is_text_content(content_type: str) -> bool:
+    base = content_type.split(";")[0].strip().lower()
+    return base in _TEXT_CONTENT_TYPES or base.startswith("text/")
+
+
+def _html_to_text(html: str) -> str:
+    h = html2text.HTML2Text()
+    h.ignore_links = False
+    h.ignore_images = True
+    h.body_width = 0
+    return h.handle(html)
+
+
+class WebFetchTool(BaseTool):
+    """Safely fetch content from a public URL using SSRF-protected HTTP."""
+
+    @property
+    def name(self) -> str:
+        return "web_fetch"
+
+    @property
+    def description(self) -> str:
+        return (
+            "Fetch the content of a public web page by URL. "
+            "Returns readable text extracted from HTML by default. "
+            "Useful for reading documentation, articles, and API responses. "
+            "Only supports HTTP/HTTPS GET requests to public URLs "
+            "(private/internal network addresses are blocked)."
+        )
+
+    @property
+    def parameters(self) -> dict[str, Any]:
+        return {
+            "type": "object",
+            "properties": {
+                "url": {
+                    "type": "string",
+                    "description": "The public HTTP/HTTPS URL to fetch.",
+                },
+                "extract_text": {
+                    "type": "boolean",
+                    "description": (
+                        "If true (default), extract readable text from HTML. "
+                        "If false, return raw content."
+                    ),
+                    "default": True,
+                },
+            },
+            "required": ["url"],
+        }
+
+    @property
+    def requires_auth(self) -> bool:
+        return False
+
+    async def _execute(
+        self,
+        user_id: str | None,
+        session: ChatSession,
+        **kwargs: Any,
+    ) -> ToolResponseBase:
+        url: str = (kwargs.get("url") or "").strip()
+        extract_text: bool = kwargs.get("extract_text", True)
+        session_id = session.session_id if session else None
+
+        if not url:
+            return ErrorResponse(
+                message="Please provide a URL to fetch.",
+                error="missing_url",
+                session_id=session_id,
+            )
+
+        try:
+            client = Requests(raise_for_status=False, retry_max_attempts=1)
+            response = await client.get(url, timeout=_REQUEST_TIMEOUT)
+        except ValueError as e:
+            # validate_url raises ValueError for SSRF / blocked IPs
+            return ErrorResponse(
+                message=f"URL blocked: {e}",
+                error="url_blocked",
+                session_id=session_id,
+            )
+        except Exception as e:
+            logger.warning(f"[web_fetch] Request failed for {url}: {e}")
+            return ErrorResponse(
+                message=f"Failed to fetch URL: {e}",
+                error="fetch_failed",
+                session_id=session_id,
+            )
+
+        content_type = response.headers.get("content-type", "")
+        if not _is_text_content(content_type):
+            return ErrorResponse(
+                message=f"Non-text content type: {content_type.split(';')[0]}",
+                error="unsupported_content_type",
+                session_id=session_id,
+            )
+
+        raw = response.content[:_MAX_CONTENT_BYTES]
+        text = raw.decode("utf-8", errors="replace")
+
+        if extract_text and "html" in content_type.lower():
+            text = _html_to_text(text)
+
+        truncated = len(text) > _MAX_OUTPUT_CHARS
+        if truncated:
+            text = text[:_MAX_OUTPUT_CHARS]
+
+        return WebFetchResponse(
+            message=f"Fetched {url}" + (" (truncated)" if truncated else ""),
+            url=response.url,
+            status_code=response.status,
+            content_type=content_type.split(";")[0].strip(),
+            content=text,
+            truncated=truncated,
+            session_id=session_id,
+        )

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

@@ -88,7 +88,9 @@ class ListWorkspaceFilesTool(BaseTool):
     @property
     def description(self) -> str:
         return (
-            "List files in the user's workspace. "
+            "List files in the user's persistent workspace (cloud storage). "
+            "These files survive across sessions. "
+            "For ephemeral session files, use the SDK Read/Glob tools instead. "
             "Returns file names, paths, sizes, and metadata. "
             "Optionally filter by path prefix."
         )
@@ -204,7 +206,9 @@ class ReadWorkspaceFileTool(BaseTool):
     @property
     def description(self) -> str:
         return (
-            "Read a file from the user's workspace. "
+            "Read a file from the user's persistent workspace (cloud storage). "
+            "These files survive across sessions. "
+            "For ephemeral session files, use the SDK Read tool instead. "
             "Specify either file_id or path to identify the file. "
             "For small text files, returns content directly. "
             "For large or binary files, returns metadata and a download URL. "
@@ -378,7 +382,9 @@ class WriteWorkspaceFileTool(BaseTool):
     @property
     def description(self) -> str:
         return (
-            "Write or create a file in the user's workspace. "
+            "Write or create a file in the user's persistent workspace (cloud storage). "
+            "These files survive across sessions. "
+            "For ephemeral session files, use the SDK Write tool instead. "
             "Provide the content as a base64-encoded string. "
             f"Maximum file size is {Config().max_file_size_mb}MB. "
             "Files are saved to the current session's folder by default. "
@@ -523,7 +529,7 @@ class DeleteWorkspaceFileTool(BaseTool):
     @property
     def description(self) -> str:
         return (
-            "Delete a file from the user's workspace. "
+            "Delete a file from the user's persistent workspace (cloud storage). "
             "Specify either file_id or path to identify the file. "
             "Paths are scoped to the current session by default. "
             "Use /sessions/<session_id>/... for cross-session access."

autogpt_platform/backend/backend/blocks/claude_code.py

@@ -1,10 +1,10 @@
 import json
 import shlex
 import uuid
-from typing import Literal, Optional
+from typing import TYPE_CHECKING, Literal, Optional
 
 from e2b import AsyncSandbox as BaseAsyncSandbox
-from pydantic import BaseModel, SecretStr
+from pydantic import SecretStr
 
 from backend.blocks._base import (
     Block,
@@ -20,6 +20,13 @@ from backend.data.model import (
     SchemaField,
 )
 from backend.integrations.providers import ProviderName
+from backend.util.sandbox_files import (
+    SandboxFileOutput,
+    extract_and_store_sandbox_files,
+)
+
+if TYPE_CHECKING:
+    from backend.executor.utils import ExecutionContext
 
 
 class ClaudeCodeExecutionError(Exception):
@@ -174,22 +181,15 @@ class ClaudeCodeBlock(Block):
             advanced=True,
         )
 
-    class FileOutput(BaseModel):
-        """A file extracted from the sandbox."""
-
-        path: str
-        relative_path: str  # Path relative to working directory (for GitHub, etc.)
-        name: str
-        content: str
-
     class Output(BlockSchemaOutput):
         response: str = SchemaField(
             description="The output/response from Claude Code execution"
         )
-        files: list["ClaudeCodeBlock.FileOutput"] = SchemaField(
+        files: list[SandboxFileOutput] = SchemaField(
             description=(
                 "List of text files created/modified by Claude Code during this execution. "
-                "Each file has 'path', 'relative_path', 'name', and 'content' fields."
+                "Each file has 'path', 'relative_path', 'name', 'content', and 'workspace_ref' fields. "
+                "workspace_ref contains a workspace:// URI if the file was stored to workspace."
             )
         )
         conversation_history: str = SchemaField(
@@ -252,6 +252,7 @@ class ClaudeCodeBlock(Block):
                             "relative_path": "index.html",
                             "name": "index.html",
                             "content": "<html>Hello World</html>",
+                            "workspace_ref": None,
                         }
                     ],
                 ),
@@ -267,11 +268,12 @@ class ClaudeCodeBlock(Block):
                 "execute_claude_code": lambda *args, **kwargs: (
                     "Created index.html with hello world content",  # response
                     [
-                        ClaudeCodeBlock.FileOutput(
+                        SandboxFileOutput(
                             path="/home/user/index.html",
                             relative_path="index.html",
                             name="index.html",
                             content="<html>Hello World</html>",
+                            workspace_ref=None,
                         )
                     ],  # files
                     "User: Create a hello world HTML file\n"
@@ -294,7 +296,8 @@ class ClaudeCodeBlock(Block):
         existing_sandbox_id: str,
         conversation_history: str,
         dispose_sandbox: bool,
-    ) -> tuple[str, list["ClaudeCodeBlock.FileOutput"], str, str, str]:
+        execution_context: "ExecutionContext",
+    ) -> tuple[str, list[SandboxFileOutput], str, str, str]:
         """
         Execute Claude Code in an E2B sandbox.
 
@@ -449,14 +452,18 @@ class ClaudeCodeBlock(Block):
                 else:
                     new_conversation_history = turn_entry
 
-            # Extract files created/modified during this run
+            # Extract files created/modified during this run and store to workspace
-            files = await self._extract_files(
+            sandbox_files = await extract_and_store_sandbox_files(
-                sandbox, working_directory, start_timestamp
+                sandbox=sandbox,
+                working_directory=working_directory,
+                execution_context=execution_context,
+                since_timestamp=start_timestamp,
+                text_only=True,
             )
 
             return (
                 response,
-                files,
+                sandbox_files,  # Already SandboxFileOutput objects
                 new_conversation_history,
                 current_session_id,
                 sandbox_id,
@@ -471,140 +478,6 @@ class ClaudeCodeBlock(Block):
             if dispose_sandbox and sandbox:
                 await sandbox.kill()
 
-    async def _extract_files(
-        self,
-        sandbox: BaseAsyncSandbox,
-        working_directory: str,
-        since_timestamp: str | None = None,
-    ) -> list["ClaudeCodeBlock.FileOutput"]:
-        """
-        Extract text files created/modified during this Claude Code execution.
-
-        Args:
-            sandbox: The E2B sandbox instance
-            working_directory: Directory to search for files
-            since_timestamp: ISO timestamp - only return files modified after this time
-
-        Returns:
-            List of FileOutput objects with path, relative_path, name, and content
-        """
-        files: list[ClaudeCodeBlock.FileOutput] = []
-
-        # Text file extensions we can safely read as text
-        text_extensions = {
-            ".txt",
-            ".md",
-            ".html",
-            ".htm",
-            ".css",
-            ".js",
-            ".ts",
-            ".jsx",
-            ".tsx",
-            ".json",
-            ".xml",
-            ".yaml",
-            ".yml",
-            ".toml",
-            ".ini",
-            ".cfg",
-            ".conf",
-            ".py",
-            ".rb",
-            ".php",
-            ".java",
-            ".c",
-            ".cpp",
-            ".h",
-            ".hpp",
-            ".cs",
-            ".go",
-            ".rs",
-            ".swift",
-            ".kt",
-            ".scala",
-            ".sh",
-            ".bash",
-            ".zsh",
-            ".sql",
-            ".graphql",
-            ".env",
-            ".gitignore",
-            ".dockerfile",
-            "Dockerfile",
-            ".vue",
-            ".svelte",
-            ".astro",
-            ".mdx",
-            ".rst",
-            ".tex",
-            ".csv",
-            ".log",
-        }
-
-        try:
-            # List files recursively using find command
-            # Exclude node_modules and .git directories, but allow hidden files
-            # like .env and .gitignore (they're filtered by text_extensions later)
-            # Filter by timestamp to only get files created/modified during this run
-            safe_working_dir = shlex.quote(working_directory)
-            timestamp_filter = ""
-            if since_timestamp:
-                timestamp_filter = f"-newermt {shlex.quote(since_timestamp)} "
-            find_result = await sandbox.commands.run(
-                f"find {safe_working_dir} -type f "
-                f"{timestamp_filter}"
-                f"-not -path '*/node_modules/*' "
-                f"-not -path '*/.git/*' "
-                f"2>/dev/null"
-            )
-
-            if find_result.stdout:
-                for file_path in find_result.stdout.strip().split("\n"):
-                    if not file_path:
-                        continue
-
-                    # Check if it's a text file we can read
-                    is_text = any(
-                        file_path.endswith(ext) for ext in text_extensions
-                    ) or file_path.endswith("Dockerfile")
-
-                    if is_text:
-                        try:
-                            content = await sandbox.files.read(file_path)
-                            # Handle bytes or string
-                            if isinstance(content, bytes):
-                                content = content.decode("utf-8", errors="replace")
-
-                            # Extract filename from path
-                            file_name = file_path.split("/")[-1]
-
-                            # Calculate relative path by stripping working directory
-                            relative_path = file_path
-                            if file_path.startswith(working_directory):
-                                relative_path = file_path[len(working_directory) :]
-                                # Remove leading slash if present
-                                if relative_path.startswith("/"):
-                                    relative_path = relative_path[1:]
-
-                            files.append(
-                                ClaudeCodeBlock.FileOutput(
-                                    path=file_path,
-                                    relative_path=relative_path,
-                                    name=file_name,
-                                    content=content,
-                                )
-                            )
-                        except Exception:
-                            # Skip files that can't be read
-                            pass
-
-        except Exception:
-            # If file extraction fails, return empty results
-            pass
-
-        return files
-
     def _escape_prompt(self, prompt: str) -> str:
         """Escape the prompt for safe shell execution."""
         # Use single quotes and escape any single quotes in the prompt
@@ -617,6 +490,7 @@ class ClaudeCodeBlock(Block):
         *,
         e2b_credentials: APIKeyCredentials,
         anthropic_credentials: APIKeyCredentials,
+        execution_context: "ExecutionContext",
         **kwargs,
     ) -> BlockOutput:
         try:
@@ -637,6 +511,7 @@ class ClaudeCodeBlock(Block):
                 existing_sandbox_id=input_data.sandbox_id,
                 conversation_history=input_data.conversation_history,
                 dispose_sandbox=input_data.dispose_sandbox,
+                execution_context=execution_context,
             )
 
             yield "response", response

autogpt_platform/backend/backend/blocks/code_executor.py

@@ -1,5 +1,5 @@
 from enum import Enum
-from typing import Any, Literal, Optional
+from typing import TYPE_CHECKING, Any, Literal, Optional
 
 from e2b_code_interpreter import AsyncSandbox
 from e2b_code_interpreter import Result as E2BExecutionResult
@@ -20,6 +20,13 @@ from backend.data.model import (
     SchemaField,
 )
 from backend.integrations.providers import ProviderName
+from backend.util.sandbox_files import (
+    SandboxFileOutput,
+    extract_and_store_sandbox_files,
+)
+
+if TYPE_CHECKING:
+    from backend.executor.utils import ExecutionContext
 
 TEST_CREDENTIALS = APIKeyCredentials(
     id="01234567-89ab-cdef-0123-456789abcdef",
@@ -85,6 +92,9 @@ class CodeExecutionResult(MainCodeExecutionResult):
 class BaseE2BExecutorMixin:
     """Shared implementation methods for E2B executor blocks."""
 
+    # Default working directory in E2B sandboxes
+    WORKING_DIR = "/home/user"
+
     async def execute_code(
         self,
         api_key: str,
@@ -95,14 +105,21 @@ class BaseE2BExecutorMixin:
         timeout: Optional[int] = None,
         sandbox_id: Optional[str] = None,
         dispose_sandbox: bool = False,
+        execution_context: Optional["ExecutionContext"] = None,
+        extract_files: bool = False,
     ):
         """
         Unified code execution method that handles all three use cases:
         1. Create new sandbox and execute (ExecuteCodeBlock)
         2. Create new sandbox, execute, and return sandbox_id (InstantiateCodeSandboxBlock)
         3. Connect to existing sandbox and execute (ExecuteCodeStepBlock)
+
+        Args:
+            extract_files: If True and execution_context provided, extract files
+                           created/modified during execution and store to workspace.
         """  # noqa
         sandbox = None
+        files: list[SandboxFileOutput] = []
         try:
             if sandbox_id:
                 # Connect to existing sandbox (ExecuteCodeStepBlock case)
@@ -118,6 +135,12 @@ class BaseE2BExecutorMixin:
                     for cmd in setup_commands:
                         await sandbox.commands.run(cmd)
 
+            # Capture timestamp before execution to scope file extraction
+            start_timestamp = None
+            if extract_files:
+                ts_result = await sandbox.commands.run("date -u +%Y-%m-%dT%H:%M:%S")
+                start_timestamp = ts_result.stdout.strip() if ts_result.stdout else None
+
             # Execute the code
             execution = await sandbox.run_code(
                 code,
@@ -133,7 +156,24 @@ class BaseE2BExecutorMixin:
             stdout_logs = "".join(execution.logs.stdout)
             stderr_logs = "".join(execution.logs.stderr)
 
-            return results, text_output, stdout_logs, stderr_logs, sandbox.sandbox_id
+            # Extract files created/modified during this execution
+            if extract_files and execution_context:
+                files = await extract_and_store_sandbox_files(
+                    sandbox=sandbox,
+                    working_directory=self.WORKING_DIR,
+                    execution_context=execution_context,
+                    since_timestamp=start_timestamp,
+                    text_only=False,  # Include binary files too
+                )
+
+            return (
+                results,
+                text_output,
+                stdout_logs,
+                stderr_logs,
+                sandbox.sandbox_id,
+                files,
+            )
         finally:
             # Dispose of sandbox if requested to reduce usage costs
             if dispose_sandbox and sandbox:
@@ -238,6 +278,12 @@ class ExecuteCodeBlock(Block, BaseE2BExecutorMixin):
             description="Standard output logs from execution"
         )
         stderr_logs: str = SchemaField(description="Standard error logs from execution")
+        files: list[SandboxFileOutput] = SchemaField(
+            description=(
+                "Files created or modified during execution. "
+                "Each file has path, name, content, and workspace_ref (if stored)."
+            ),
+        )
 
     def __init__(self):
         super().__init__(
@@ -259,23 +305,30 @@ class ExecuteCodeBlock(Block, BaseE2BExecutorMixin):
                 ("results", []),
                 ("response", "Hello World"),
                 ("stdout_logs", "Hello World\n"),
+                ("files", []),
             ],
             test_mock={
-                "execute_code": lambda api_key, code, language, template_id, setup_commands, timeout, dispose_sandbox: (  # noqa
+                "execute_code": lambda api_key, code, language, template_id, setup_commands, timeout, dispose_sandbox, execution_context, extract_files: (  # noqa
                     [],  # results
                     "Hello World",  # text_output
                     "Hello World\n",  # stdout_logs
                     "",  # stderr_logs
                     "sandbox_id",  # sandbox_id
+                    [],  # files
                 ),
             },
         )
 
     async def run(
-        self, input_data: Input, *, credentials: APIKeyCredentials, **kwargs
+        self,
+        input_data: Input,
+        *,
+        credentials: APIKeyCredentials,
+        execution_context: "ExecutionContext",
+        **kwargs,
     ) -> BlockOutput:
         try:
-            results, text_output, stdout, stderr, _ = await self.execute_code(
+            results, text_output, stdout, stderr, _, files = await self.execute_code(
                 api_key=credentials.api_key.get_secret_value(),
                 code=input_data.code,
                 language=input_data.language,
@@ -283,6 +336,8 @@ class ExecuteCodeBlock(Block, BaseE2BExecutorMixin):
                 setup_commands=input_data.setup_commands,
                 timeout=input_data.timeout,
                 dispose_sandbox=input_data.dispose_sandbox,
+                execution_context=execution_context,
+                extract_files=True,
             )
 
             # Determine result object shape & filter out empty formats
@@ -296,6 +351,8 @@ class ExecuteCodeBlock(Block, BaseE2BExecutorMixin):
                 yield "stdout_logs", stdout
             if stderr:
                 yield "stderr_logs", stderr
+            # Always yield files (empty list if none)
+            yield "files", [f.model_dump() for f in files]
         except Exception as e:
             yield "error", str(e)
 
@@ -393,6 +450,7 @@ class InstantiateCodeSandboxBlock(Block, BaseE2BExecutorMixin):
                     "Hello World\n",  # stdout_logs
                     "",  # stderr_logs
                     "sandbox_id",  # sandbox_id
+                    [],  # files
                 ),
             },
         )
@@ -401,7 +459,7 @@ class InstantiateCodeSandboxBlock(Block, BaseE2BExecutorMixin):
         self, input_data: Input, *, credentials: APIKeyCredentials, **kwargs
     ) -> BlockOutput:
         try:
-            _, text_output, stdout, stderr, sandbox_id = await self.execute_code(
+            _, text_output, stdout, stderr, sandbox_id, _ = await self.execute_code(
                 api_key=credentials.api_key.get_secret_value(),
                 code=input_data.setup_code,
                 language=input_data.language,
@@ -500,6 +558,7 @@ class ExecuteCodeStepBlock(Block, BaseE2BExecutorMixin):
                     "Hello World\n",  # stdout_logs
                     "",  # stderr_logs
                     sandbox_id,  # sandbox_id
+                    [],  # files
                 ),
             },
         )
@@ -508,7 +567,7 @@ class ExecuteCodeStepBlock(Block, BaseE2BExecutorMixin):
         self, input_data: Input, *, credentials: APIKeyCredentials, **kwargs
     ) -> BlockOutput:
         try:
-            results, text_output, stdout, stderr, _ = await self.execute_code(
+            results, text_output, stdout, stderr, _, _ = await self.execute_code(
                 api_key=credentials.api_key.get_secret_value(),
                 code=input_data.step_code,
                 language=input_data.language,

autogpt_platform/backend/backend/util/feature_flag.py

@@ -38,6 +38,7 @@ class Flag(str, Enum):
     AGENT_ACTIVITY = "agent-activity"
     ENABLE_PLATFORM_PAYMENT = "enable-platform-payment"
     CHAT = "chat"
+    COPILOT_SDK = "copilot-sdk"
 
 
 def is_configured() -> bool:

autogpt_platform/backend/backend/util/sandbox_files.py

@@ -0,0 +1,288 @@
+"""
+Shared utilities for extracting and storing files from E2B sandboxes.
+
+This module provides common file extraction and workspace storage functionality
+for blocks that run code in E2B sandboxes (Claude Code, Code Executor, etc.).
+"""
+
+import base64
+import logging
+import mimetypes
+import shlex
+from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+from pydantic import BaseModel
+
+from backend.util.file import store_media_file
+from backend.util.type import MediaFileType
+
+if TYPE_CHECKING:
+    from e2b import AsyncSandbox as BaseAsyncSandbox
+
+    from backend.executor.utils import ExecutionContext
+
+logger = logging.getLogger(__name__)
+
+# Text file extensions that can be safely read and stored as text
+TEXT_EXTENSIONS = {
+    ".txt",
+    ".md",
+    ".html",
+    ".htm",
+    ".css",
+    ".js",
+    ".ts",
+    ".jsx",
+    ".tsx",
+    ".json",
+    ".xml",
+    ".yaml",
+    ".yml",
+    ".toml",
+    ".ini",
+    ".cfg",
+    ".conf",
+    ".py",
+    ".rb",
+    ".php",
+    ".java",
+    ".c",
+    ".cpp",
+    ".h",
+    ".hpp",
+    ".cs",
+    ".go",
+    ".rs",
+    ".swift",
+    ".kt",
+    ".scala",
+    ".sh",
+    ".bash",
+    ".zsh",
+    ".sql",
+    ".graphql",
+    ".env",
+    ".gitignore",
+    ".dockerfile",
+    "Dockerfile",
+    ".vue",
+    ".svelte",
+    ".astro",
+    ".mdx",
+    ".rst",
+    ".tex",
+    ".csv",
+    ".log",
+}
+
+
+class SandboxFileOutput(BaseModel):
+    """A file extracted from a sandbox and optionally stored in workspace."""
+
+    path: str
+    """Full path in the sandbox."""
+
+    relative_path: str
+    """Path relative to the working directory."""
+
+    name: str
+    """Filename only."""
+
+    content: str
+    """File content as text (for backward compatibility)."""
+
+    workspace_ref: str | None = None
+    """Workspace reference (workspace://{id}#mime) if stored, None otherwise."""
+
+
+@dataclass
+class ExtractedFile:
+    """Internal representation of an extracted file before storage."""
+
+    path: str
+    relative_path: str
+    name: str
+    content: bytes
+    is_text: bool
+
+
+async def extract_sandbox_files(
+    sandbox: "BaseAsyncSandbox",
+    working_directory: str,
+    since_timestamp: str | None = None,
+    text_only: bool = True,
+) -> list[ExtractedFile]:
+    """
+    Extract files from an E2B sandbox.
+
+    Args:
+        sandbox: The E2B sandbox instance
+        working_directory: Directory to search for files
+        since_timestamp: ISO timestamp - only return files modified after this time
+        text_only: If True, only extract text files (default). If False, extract all files.
+
+    Returns:
+        List of ExtractedFile objects with path, content, and metadata
+    """
+    files: list[ExtractedFile] = []
+
+    try:
+        # Build find command
+        safe_working_dir = shlex.quote(working_directory)
+        timestamp_filter = ""
+        if since_timestamp:
+            timestamp_filter = f"-newermt {shlex.quote(since_timestamp)} "
+
+        find_result = await sandbox.commands.run(
+            f"find {safe_working_dir} -type f "
+            f"{timestamp_filter}"
+            f"-not -path '*/node_modules/*' "
+            f"-not -path '*/.git/*' "
+            f"2>/dev/null"
+        )
+
+        if not find_result.stdout:
+            return files
+
+        for file_path in find_result.stdout.strip().split("\n"):
+            if not file_path:
+                continue
+
+            # Check if it's a text file
+            is_text = any(file_path.endswith(ext) for ext in TEXT_EXTENSIONS)
+
+            # Skip non-text files if text_only mode
+            if text_only and not is_text:
+                continue
+
+            try:
+                # Read file content as bytes
+                content = await sandbox.files.read(file_path, format="bytes")
+                if isinstance(content, str):
+                    content = content.encode("utf-8")
+                elif isinstance(content, bytearray):
+                    content = bytes(content)
+
+                # Extract filename from path
+                file_name = file_path.split("/")[-1]
+
+                # Calculate relative path
+                relative_path = file_path
+                if file_path.startswith(working_directory):
+                    relative_path = file_path[len(working_directory) :]
+                    if relative_path.startswith("/"):
+                        relative_path = relative_path[1:]
+
+                files.append(
+                    ExtractedFile(
+                        path=file_path,
+                        relative_path=relative_path,
+                        name=file_name,
+                        content=content,
+                        is_text=is_text,
+                    )
+                )
+            except Exception as e:
+                logger.debug(f"Failed to read file {file_path}: {e}")
+                continue
+
+    except Exception as e:
+        logger.warning(f"File extraction failed: {e}")
+
+    return files
+
+
+async def store_sandbox_files(
+    extracted_files: list[ExtractedFile],
+    execution_context: "ExecutionContext",
+) -> list[SandboxFileOutput]:
+    """
+    Store extracted sandbox files to workspace and return output objects.
+
+    Args:
+        extracted_files: List of files extracted from sandbox
+        execution_context: Execution context for workspace storage
+
+    Returns:
+        List of SandboxFileOutput objects with workspace refs
+    """
+    outputs: list[SandboxFileOutput] = []
+
+    for file in extracted_files:
+        # Decode content for text files (for backward compat content field)
+        if file.is_text:
+            try:
+                content_str = file.content.decode("utf-8", errors="replace")
+            except Exception:
+                content_str = ""
+        else:
+            content_str = f"[Binary file: {len(file.content)} bytes]"
+
+        # Build data URI (needed for storage and as binary fallback)
+        mime_type = mimetypes.guess_type(file.name)[0] or "application/octet-stream"
+        data_uri = f"data:{mime_type};base64,{base64.b64encode(file.content).decode()}"
+
+        # Try to store in workspace
+        workspace_ref: str | None = None
+        try:
+            result = await store_media_file(
+                file=MediaFileType(data_uri),
+                execution_context=execution_context,
+                return_format="for_block_output",
+            )
+            if result.startswith("workspace://"):
+                workspace_ref = result
+            elif not file.is_text:
+                # Non-workspace context (graph execution): store_media_file
+                # returned a data URI — use it as content so binary data isn't lost.
+                content_str = result
+        except Exception as e:
+            logger.warning(f"Failed to store file {file.name} to workspace: {e}")
+            # For binary files, fall back to data URI to prevent data loss
+            if not file.is_text:
+                content_str = data_uri
+
+        outputs.append(
+            SandboxFileOutput(
+                path=file.path,
+                relative_path=file.relative_path,
+                name=file.name,
+                content=content_str,
+                workspace_ref=workspace_ref,
+            )
+        )
+
+    return outputs
+
+
+async def extract_and_store_sandbox_files(
+    sandbox: "BaseAsyncSandbox",
+    working_directory: str,
+    execution_context: "ExecutionContext",
+    since_timestamp: str | None = None,
+    text_only: bool = True,
+) -> list[SandboxFileOutput]:
+    """
+    Extract files from sandbox and store them in workspace.
+
+    This is the main entry point combining extraction and storage.
+
+    Args:
+        sandbox: The E2B sandbox instance
+        working_directory: Directory to search for files
+        execution_context: Execution context for workspace storage
+        since_timestamp: ISO timestamp - only return files modified after this time
+        text_only: If True, only extract text files
+
+    Returns:
+        List of SandboxFileOutput objects with content and workspace refs
+    """
+    extracted = await extract_sandbox_files(
+        sandbox=sandbox,
+        working_directory=working_directory,
+        since_timestamp=since_timestamp,
+        text_only=text_only,
+    )
+
+    return await store_sandbox_files(extracted, execution_context)

autogpt_platform/backend/poetry.lock

@@ -897,6 +897,29 @@ files = [
     {file = "charset_normalizer-3.4.4.tar.gz", hash = "sha256:94537985111c35f28720e43603b8e7b43a6ecfb2ce1d3058bbe955b73404e21a"},
 ]
 
+[[package]]
+name = "claude-agent-sdk"
+version = "0.1.35"
+description = "Python SDK for Claude Code"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "claude_agent_sdk-0.1.35-py3-none-macosx_11_0_arm64.whl", hash = "sha256:df67f4deade77b16a9678b3a626c176498e40417f33b04beda9628287f375591"},
+    {file = "claude_agent_sdk-0.1.35-py3-none-manylinux_2_17_aarch64.whl", hash = "sha256:14963944f55ded7c8ed518feebfa5b4284aa6dd8d81aeff2e5b21a962ce65097"},
+    {file = "claude_agent_sdk-0.1.35-py3-none-manylinux_2_17_x86_64.whl", hash = "sha256:84344dcc535d179c1fc8a11c6f34c37c3b583447bdf09d869effb26514fd7a65"},
+    {file = "claude_agent_sdk-0.1.35-py3-none-win_amd64.whl", hash = "sha256:1b3d54b47448c93f6f372acd4d1757f047c3c1e8ef5804be7a1e3e53e2c79a5f"},
+    {file = "claude_agent_sdk-0.1.35.tar.gz", hash = "sha256:0f98e2b3c71ca85abfc042e7a35c648df88e87fda41c52e6779ef7b038dcbb52"},
+]
+
+[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"
@@ -2593,6 +2616,18 @@ 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 = "1.4.1"
@@ -3310,6 +3345,39 @@ 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"
@@ -5994,7 +6062,7 @@ 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"},
@@ -6974,6 +7042,28 @@ postgresql-psycopgbinary = ["psycopg[binary] (>=3.0.7)"]
 pymysql = ["pymysql"]
 sqlcipher = ["sqlcipher3_binary"]
 
+[[package]]
+name = "sse-starlette"
+version = "3.2.0"
+description = "SSE plugin for Starlette"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "sse_starlette-3.2.0-py3-none-any.whl", hash = "sha256:5876954bd51920fc2cd51baee47a080eb88a37b5b784e615abb0b283f801cdbf"},
+    {file = "sse_starlette-3.2.0.tar.gz", hash = "sha256:8127594edfb51abe44eac9c49e59b0b01f1039d0c7461c6fd91d4e03b70da422"},
+]
+
+[package.dependencies]
+anyio = ">=4.7.0"
+starlette = ">=0.49.1"
+
+[package.extras]
+daphne = ["daphne (>=4.2.0)"]
+examples = ["aiosqlite (>=0.21.0)", "fastapi (>=0.115.12)", "sqlalchemy[asyncio] (>=2.0.41)", "uvicorn (>=0.34.0)"]
+granian = ["granian (>=2.3.1)"]
+uvicorn = ["uvicorn (>=0.34.0)"]
+
 [[package]]
 name = "stagehand"
 version = "0.5.9"
@@ -8440,4 +8530,4 @@ cffi = ["cffi (>=1.17,<2.0) ; platform_python_implementation != \"PyPy\" and pyt
 [metadata]
 lock-version = "2.1"
 python-versions = ">=3.10,<3.14"
-content-hash = "c06e96ad49388ba7a46786e9ea55ea2c1a57408e15613237b4bee40a592a12af"
+content-hash = "942dea6daf671c3be65a22f3445feda26c1af9409d7173765e9a0742f0aa05dc"

autogpt_platform/backend/pyproject.toml

@@ -16,6 +16,7 @@ anthropic = "^0.79.0"
 apscheduler = "^3.11.1"
 autogpt-libs = { path = "../autogpt_libs", develop = true }
 bleach = { extras = ["css"], version = "^6.2.0" }
+claude-agent-sdk = "^0.1.0"
 click = "^8.2.0"
 cryptography = "^46.0"
 discord-py = "^2.5.2"

autogpt_platform/backend/test/chat/test_security_hooks.py

@@ -0,0 +1,133 @@
+"""Tests for SDK security hooks — workspace paths, tool access, and deny messages.
+
+These are pure unit tests with no external dependencies (no SDK, no DB, no server).
+They validate that the security hooks correctly block unauthorized paths,
+tool access, and dangerous input patterns.
+
+Note: Bash command validation was removed — the SDK built-in Bash tool is not in
+allowed_tools, and the bash_exec MCP tool has kernel-level network isolation
+(unshare --net) making command-level parsing unnecessary.
+"""
+
+from backend.api.features.chat.sdk.security_hooks import (
+    _validate_tool_access,
+    _validate_workspace_path,
+)
+
+SDK_CWD = "/tmp/copilot-test-session"
+
+
+def _is_denied(result: dict) -> bool:
+    hook = result.get("hookSpecificOutput", {})
+    return hook.get("permissionDecision") == "deny"
+
+
+def _reason(result: dict) -> str:
+    return result.get("hookSpecificOutput", {}).get("permissionDecisionReason", "")
+
+
+# ============================================================
+# Workspace path validation (Read, Write, Edit, etc.)
+# ============================================================
+
+
+class TestWorkspacePathValidation:
+    def test_path_in_workspace(self):
+        result = _validate_workspace_path(
+            "Read", {"file_path": f"{SDK_CWD}/file.txt"}, SDK_CWD
+        )
+        assert not _is_denied(result)
+
+    def test_path_outside_workspace(self):
+        result = _validate_workspace_path("Read", {"file_path": "/etc/passwd"}, SDK_CWD)
+        assert _is_denied(result)
+
+    def test_tool_results_allowed(self):
+        result = _validate_workspace_path(
+            "Read",
+            {"file_path": "~/.claude/projects/abc/tool-results/out.txt"},
+            SDK_CWD,
+        )
+        assert not _is_denied(result)
+
+    def test_claude_settings_blocked(self):
+        result = _validate_workspace_path(
+            "Read", {"file_path": "~/.claude/settings.json"}, SDK_CWD
+        )
+        assert _is_denied(result)
+
+    def test_claude_projects_without_tool_results(self):
+        result = _validate_workspace_path(
+            "Read", {"file_path": "~/.claude/projects/abc/credentials.json"}, SDK_CWD
+        )
+        assert _is_denied(result)
+
+    def test_no_path_allowed(self):
+        """Glob/Grep without path defaults to cwd — should be allowed."""
+        result = _validate_workspace_path("Grep", {"pattern": "foo"}, SDK_CWD)
+        assert not _is_denied(result)
+
+    def test_path_traversal_with_dotdot(self):
+        result = _validate_workspace_path(
+            "Read", {"file_path": f"{SDK_CWD}/../../../etc/passwd"}, SDK_CWD
+        )
+        assert _is_denied(result)
+
+
+# ============================================================
+# Tool access validation
+# ============================================================
+
+
+class TestToolAccessValidation:
+    def test_blocked_tools(self):
+        for tool in ("bash", "shell", "exec", "terminal", "command"):
+            result = _validate_tool_access(tool, {})
+            assert _is_denied(result), f"Tool '{tool}' should be blocked"
+
+    def test_bash_builtin_blocked(self):
+        """SDK built-in Bash (capital) is blocked as defence-in-depth."""
+        result = _validate_tool_access("Bash", {"command": "echo hello"}, SDK_CWD)
+        assert _is_denied(result)
+        assert "Bash" in _reason(result)
+
+    def test_workspace_tools_delegate(self):
+        result = _validate_tool_access(
+            "Read", {"file_path": f"{SDK_CWD}/file.txt"}, SDK_CWD
+        )
+        assert not _is_denied(result)
+
+    def test_dangerous_pattern_blocked(self):
+        result = _validate_tool_access("SomeUnknownTool", {"data": "sudo rm -rf /"})
+        assert _is_denied(result)
+
+    def test_safe_unknown_tool_allowed(self):
+        result = _validate_tool_access("SomeSafeTool", {"data": "hello world"})
+        assert not _is_denied(result)
+
+
+# ============================================================
+# Deny message quality (ntindle feedback)
+# ============================================================
+
+
+class TestDenyMessageClarity:
+    """Deny messages must include [SECURITY] and 'cannot be bypassed'
+    so the model knows the restriction is enforced, not a suggestion."""
+
+    def test_blocked_tool_message(self):
+        reason = _reason(_validate_tool_access("bash", {}))
+        assert "[SECURITY]" in reason
+        assert "cannot be bypassed" in reason
+
+    def test_bash_builtin_blocked_message(self):
+        reason = _reason(_validate_tool_access("Bash", {"command": "echo hello"}))
+        assert "[SECURITY]" in reason
+        assert "cannot be bypassed" in reason
+
+    def test_workspace_path_message(self):
+        reason = _reason(
+            _validate_workspace_path("Read", {"file_path": "/etc/passwd"}, SDK_CWD)
+        )
+        assert "[SECURITY]" in reason
+        assert "cannot be bypassed" in reason

autogpt_platform/frontend/src/app/(platform)/copilot/components/ChatMessagesContainer/ChatMessagesContainer.tsx

@@ -20,6 +20,7 @@ import { FindBlocksTool } from "../../tools/FindBlocks/FindBlocks";
 import { RunAgentTool } from "../../tools/RunAgent/RunAgent";
 import { RunBlockTool } from "../../tools/RunBlock/RunBlock";
 import { SearchDocsTool } from "../../tools/SearchDocs/SearchDocs";
+import { GenericTool } from "../../tools/GenericTool/GenericTool";
 import { ViewAgentOutputTool } from "../../tools/ViewAgentOutput/ViewAgentOutput";
 
 // ---------------------------------------------------------------------------
@@ -159,7 +160,7 @@ export const ChatMessagesContainer = ({
 
   return (
     <Conversation className="min-h-0 flex-1">
-      <ConversationContent className="flex min-h-screen flex-1 flex-col gap-6 px-3 py-6">
+      <ConversationContent className="flex flex-1 flex-col gap-6 px-3 py-6">
         {isLoading && messages.length === 0 && (
           <div className="flex min-h-full flex-1 items-center justify-center">
             <LoadingSpinner className="text-neutral-600" />
@@ -255,6 +256,16 @@ export const ChatMessagesContainer = ({
                         />
                       );
                     default:
+                      // Render a generic tool indicator for SDK built-in
+                      // tools (Read, Glob, Grep, etc.) or any unrecognized tool
+                      if (part.type.startsWith("tool-")) {
+                        return (
+                          <GenericTool
+                            key={`${message.id}-${i}`}
+                            part={part as ToolUIPart}
+                          />
+                        );
+                      }
                       return null;
                   }
                 })}

autogpt_platform/frontend/src/app/(platform)/copilot/tools/CreateAgent/CreateAgent.tsx

@@ -152,7 +152,7 @@ export function CreateAgentTool({ part }: Props) {
         <ToolAccordion {...getAccordionMeta(output)}>
           {isOperating && (
             <ContentGrid>
-              <ProgressBar value={progress} />
+              <ProgressBar value={progress} className="max-w-[280px]" />
               <ContentHint>
                 This could take a few minutes, grab a coffee ☕
               </ContentHint>

autogpt_platform/frontend/src/app/(platform)/copilot/tools/GenericTool/GenericTool.tsx

@@ -0,0 +1,63 @@
+"use client";
+
+import { ToolUIPart } from "ai";
+import { GearIcon } from "@phosphor-icons/react";
+import { MorphingTextAnimation } from "../../components/MorphingTextAnimation/MorphingTextAnimation";
+
+interface Props {
+  part: ToolUIPart;
+}
+
+function extractToolName(part: ToolUIPart): string {
+  // ToolUIPart.type is "tool-{name}", extract the name portion.
+  return part.type.replace(/^tool-/, "");
+}
+
+function formatToolName(name: string): string {
+  // "search_docs" → "Search docs", "Read" → "Read"
+  return name.replace(/_/g, " ").replace(/^\w/, (c) => c.toUpperCase());
+}
+
+function getAnimationText(part: ToolUIPart): string {
+  const label = formatToolName(extractToolName(part));
+
+  switch (part.state) {
+    case "input-streaming":
+    case "input-available":
+      return `Running ${label}…`;
+    case "output-available":
+      return `${label} completed`;
+    case "output-error":
+      return `${label} failed`;
+    default:
+      return `Running ${label}…`;
+  }
+}
+
+export function GenericTool({ part }: Props) {
+  const isStreaming =
+    part.state === "input-streaming" || part.state === "input-available";
+  const isError = part.state === "output-error";
+
+  return (
+    <div className="py-2">
+      <div className="flex items-center gap-2 text-sm text-muted-foreground">
+        <GearIcon
+          size={14}
+          weight="regular"
+          className={
+            isError
+              ? "text-red-500"
+              : isStreaming
+                ? "animate-spin text-neutral-500"
+                : "text-neutral-400"
+          }
+        />
+        <MorphingTextAnimation
+          text={getAnimationText(part)}
+          className={isError ? "text-red-500" : undefined}
+        />
+      </div>
+    </div>
+  );
+}

autogpt_platform/frontend/src/app/api/openapi.json

@@ -7022,29 +7022,24 @@
           "input_schema": {
             "additionalProperties": true,
             "type": "object",
-            "title": "Input Schema"
+            "title": "Input Schema",
+            "description": "Full JSON schema for block inputs"
           },
           "output_schema": {
             "additionalProperties": true,
             "type": "object",
-            "title": "Output Schema"
+            "title": "Output Schema",
+            "description": "Full JSON schema for block outputs"
           },
           "required_inputs": {
             "items": { "$ref": "#/components/schemas/BlockInputFieldInfo" },
             "type": "array",
             "title": "Required Inputs",
-            "description": "List of required input fields for this block"
+            "description": "List of input fields for this block"
           }
         },
         "type": "object",
-        "required": [
+        "required": ["id", "name", "description", "categories"],
-          "id",
-          "name",
-          "description",
-          "categories",
-          "input_schema",
-          "output_schema"
-        ],
         "title": "BlockInfoSummary",
         "description": "Summary of a block for search results."
       },
@@ -7090,7 +7085,7 @@
           "usage_hint": {
             "type": "string",
             "title": "Usage Hint",
-            "default": "To execute a block, call run_block with block_id set to the block's 'id' field and input_data containing the required fields from input_schema."
+            "default": "To execute a block, call run_block with block_id set to the block's 'id' field and input_data containing the fields listed in required_inputs."
           }
         },
         "type": "object",
@@ -10495,7 +10490,10 @@
           "operation_started",
           "operation_pending",
           "operation_in_progress",
-          "input_validation_error"
+          "input_validation_error",
+          "web_fetch",
+          "bash_exec",
+          "operation_status"
         ],
         "title": "ResponseType",
         "description": "Types of tool responses."

autogpt_platform/frontend/src/components/contextual/CredentialsInput/components/APIKeyCredentialsModal/APIKeyCredentialsModal.tsx

@@ -30,6 +30,7 @@ export function APIKeyCredentialsModal({
   const {
     form,
     isLoading,
+    isSubmitting,
     supportsApiKey,
     providerName,
     schemaDescription,
@@ -138,7 +139,12 @@ export function APIKeyCredentialsModal({
                 />
               )}
             />
-            <Button type="submit" className="min-w-68">
+            <Button
+              type="submit"
+              className="min-w-68"
+              loading={isSubmitting}
+              disabled={isSubmitting}
+            >
               Add API Key
             </Button>
           </form>

autogpt_platform/frontend/src/components/contextual/CredentialsInput/components/APIKeyCredentialsModal/useAPIKeyCredentialsModal.ts

@@ -4,6 +4,7 @@ import {
   CredentialsMetaInput,
 } from "@/lib/autogpt-server-api/types";
 import { zodResolver } from "@hookform/resolvers/zod";
+import { useState } from "react";
 import { useForm, type UseFormReturn } from "react-hook-form";
 import { z } from "zod";
 
@@ -26,6 +27,7 @@ export function useAPIKeyCredentialsModal({
 }: Args): {
   form: UseFormReturn<APIKeyFormValues>;
   isLoading: boolean;
+  isSubmitting: boolean;
   supportsApiKey: boolean;
   provider?: string;
   providerName?: string;
@@ -33,6 +35,7 @@ export function useAPIKeyCredentialsModal({
   onSubmit: (values: APIKeyFormValues) => Promise<void>;
 } {
   const credentials = useCredentials(schema, siblingInputs);
+  const [isSubmitting, setIsSubmitting] = useState(false);
 
   const formSchema = z.object({
     apiKey: z.string().min(1, "API Key is required"),
@@ -40,48 +43,42 @@ export function useAPIKeyCredentialsModal({
     expiresAt: z.string().optional(),
   });
 
-  function getDefaultExpirationDate(): string {
-    const tomorrow = new Date();
-    tomorrow.setDate(tomorrow.getDate() + 1);
-    tomorrow.setHours(0, 0, 0, 0);
-    const year = tomorrow.getFullYear();
-    const month = String(tomorrow.getMonth() + 1).padStart(2, "0");
-    const day = String(tomorrow.getDate()).padStart(2, "0");
-    const hours = String(tomorrow.getHours()).padStart(2, "0");
-    const minutes = String(tomorrow.getMinutes()).padStart(2, "0");
-    return `${year}-${month}-${day}T${hours}:${minutes}`;
-  }
-
   const form = useForm<APIKeyFormValues>({
     resolver: zodResolver(formSchema),
     defaultValues: {
       apiKey: "",
       title: "",
-      expiresAt: getDefaultExpirationDate(),
+      expiresAt: "",
     },
   });
 
   async function onSubmit(values: APIKeyFormValues) {
     if (!credentials || credentials.isLoading) return;
-    const expiresAt = values.expiresAt
+    setIsSubmitting(true);
-      ? new Date(values.expiresAt).getTime() / 1000
+    try {
-      : undefined;
+      const expiresAt = values.expiresAt
-    const newCredentials = await credentials.createAPIKeyCredentials({
+        ? new Date(values.expiresAt).getTime() / 1000
-      api_key: values.apiKey,
+        : undefined;
-      title: values.title,
+      const newCredentials = await credentials.createAPIKeyCredentials({
-      expires_at: expiresAt,
+        api_key: values.apiKey,
-    });
+        title: values.title,
-    onCredentialsCreate({
+        expires_at: expiresAt,
-      provider: credentials.provider,
+      });
-      id: newCredentials.id,
+      onCredentialsCreate({
-      type: "api_key",
+        provider: credentials.provider,
-      title: newCredentials.title,
+        id: newCredentials.id,
-    });
+        type: "api_key",
+        title: newCredentials.title,
+      });
+    } finally {
+      setIsSubmitting(false);
+    }
   }
 
   return {
     form,
     isLoading: !credentials || credentials.isLoading,
+    isSubmitting,
     supportsApiKey: !!credentials?.supportsApiKey,
     provider: credentials?.provider,
     providerName:

docs/integrations/block-integrations/llm.md

@@ -563,7 +563,7 @@ The block supports conversation continuation through three mechanisms:
 |--------|-------------|------|
 | error | Error message if execution failed | str |
 | response | The output/response from Claude Code execution | str |
-| files | List of text files created/modified by Claude Code during this execution. Each file has 'path', 'relative_path', 'name', and 'content' fields. | List[FileOutput] |
+| files | List of text files created/modified by Claude Code during this execution. Each file has 'path', 'relative_path', 'name', 'content', and 'workspace_ref' fields. workspace_ref contains a workspace:// URI if the file was stored to workspace. | List[SandboxFileOutput] |
 | conversation_history | Full conversation history including this turn. Pass this to conversation_history input to continue on a fresh sandbox if the previous sandbox timed out. | str |
 | session_id | Session ID for this conversation. Pass this back along with sandbox_id to continue the conversation. | str |
 | sandbox_id | ID of the sandbox instance. Pass this back along with session_id to continue the conversation. This is None if dispose_sandbox was True (sandbox was disposed). | str |

docs/integrations/block-integrations/misc.md

@@ -215,6 +215,7 @@ The sandbox includes pip and npm pre-installed. Set timeout to limit execution t
 | response | Text output (if any) of the main execution result | str |
 | stdout_logs | Standard output logs from execution | str |
 | stderr_logs | Standard error logs from execution | str |
+| files | Files created or modified during execution. Each file has path, name, content, and workspace_ref (if stored). | List[SandboxFileOutput] |
 
 ### Possible use case
 <!-- MANUAL: use_case -->