fix(copilot): RPC DataError reconstruction, chat stream reconnection

Fix two issues:

1. RPC DataError deserialization crash: When the database-manager
   returns a 400 for a Prisma DataError/UniqueViolationError, the
   client-side reconstruction crashes because DataError.__init__
   expects a dict but exc.args only contains a string message.
   Wrap the string in the expected dict structure so the exception
   is properly caught by callers (e.g. workspace file overwrites).

2. Chat stream reconnection on page refresh: The AI SDK's built-in
   resume:true fires before message hydration completes, causing
   hydrated messages to overwrite the resumed stream. Replace with
   manual resumeStream() called after hydration + active_stream
   detection. Show the stop button immediately when an active stream
   is detected (isReconnecting flag) and prevent sending new messages
   until reconnected.

autogpt_platform/backend/backend/copilot/executor/manager.py

@@ -4,7 +4,6 @@ This module contains the CoPilotExecutor class that consumes chat tasks from
 RabbitMQ and processes them using a thread pool, following the graph executor pattern.
 """
 
-import asyncio
 import logging
 import os
 import threading
@@ -164,21 +163,23 @@ class CoPilotExecutor(AppProcess):
                 self._cancel_thread, self.cancel_client, "[cleanup][cancel]"
             )
 
-        # Shutdown executor
+        # Clean up worker threads (closes per-loop workspace storage sessions)
         if self._executor:
+            from .processor import cleanup_worker
+
+            logger.info(f"[cleanup {pid}] Cleaning up workers...")
+            futures = []
+            for _ in range(self._executor._max_workers):
+                futures.append(self._executor.submit(cleanup_worker))
+            for f in futures:
+                try:
+                    f.result(timeout=10)
+                except Exception as e:
+                    logger.warning(f"[cleanup {pid}] Worker cleanup error: {e}")
+
             logger.info(f"[cleanup {pid}] Shutting down executor...")
             self._executor.shutdown(wait=False)
 
-        # Close async resources (workspace storage aiohttp session, etc.)
-        try:
-            from backend.util.workspace_storage import shutdown_workspace_storage
-
-            loop = asyncio.new_event_loop()
-            loop.run_until_complete(shutdown_workspace_storage())
-            loop.close()
-        except Exception as e:
-            logger.warning(f"[cleanup {pid}] Error closing workspace storage: {e}")
-
         # Release any remaining locks
         for task_id, lock in list(self._task_locks.items()):
             try:

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

@@ -60,6 +60,18 @@ def init_worker():
     _tls.processor.on_executor_start()
 
 
+def cleanup_worker():
+    """Clean up the processor for the current worker thread.
+
+    Should be called before the worker thread's event loop is destroyed so
+    that event-loop-bound resources (e.g. ``aiohttp.ClientSession``) are
+    closed on the correct loop.
+    """
+    processor: CoPilotProcessor | None = getattr(_tls, "processor", None)
+    if processor is not None:
+        processor.cleanup()
+
+
 # ============ Processor Class ============ #
 
 
@@ -98,6 +110,28 @@ class CoPilotProcessor:
 
         logger.info(f"[CoPilotExecutor] Worker {self.tid} started")
 
+    def cleanup(self):
+        """Clean up event-loop-bound resources before the loop is destroyed.
+
+        Shuts down the workspace storage instance that belongs to this
+        worker's event loop, ensuring ``aiohttp.ClientSession.close()``
+        runs on the same loop that created the session.
+        """
+        from backend.util.workspace_storage import shutdown_workspace_storage
+
+        try:
+            future = asyncio.run_coroutine_threadsafe(
+                shutdown_workspace_storage(), self.execution_loop
+            )
+            future.result(timeout=5)
+        except Exception as e:
+            logger.warning(f"[CoPilotExecutor] Worker {self.tid} cleanup error: {e}")
+
+        # Stop the event loop
+        self.execution_loop.call_soon_threadsafe(self.execution_loop.stop)
+        self.execution_thread.join(timeout=5)
+        logger.info(f"[CoPilotExecutor] Worker {self.tid} cleaned up")
+
     @error_logged(swallow=False)
     def execute(
         self,

autogpt_platform/backend/backend/copilot/sdk/response_adapter.py

@@ -53,6 +53,7 @@ class SDKResponseAdapter:
         self.has_started_text = False
         self.has_ended_text = False
         self.current_tool_calls: dict[str, dict[str, str]] = {}
+        self.resolved_tool_calls: set[str] = set()
         self.task_id: str | None = None
         self.step_open = False
 
@@ -74,6 +75,10 @@ class SDKResponseAdapter:
                 self.step_open = True
 
         elif isinstance(sdk_message, AssistantMessage):
+            # Flush any SDK built-in tool calls that didn't get a UserMessage
+            # result (e.g. WebSearch, Read handled internally by the CLI).
+            self._flush_unresolved_tool_calls(responses)
+
             # 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:
@@ -111,6 +116,8 @@ class SDKResponseAdapter:
             # UserMessage carries tool results back from tool execution.
             content = sdk_message.content
             blocks = content if isinstance(content, list) else []
+            resolved_in_blocks: set[str] = set()
+
             for block in blocks:
                 if isinstance(block, ToolResultBlock) and block.tool_use_id:
                     tool_info = self.current_tool_calls.get(block.tool_use_id, {})
@@ -132,6 +139,37 @@ class SDKResponseAdapter:
                             success=not (block.is_error or False),
                         )
                     )
+                    resolved_in_blocks.add(block.tool_use_id)
+
+            # Handle SDK built-in tool results carried via parent_tool_use_id
+            # instead of (or in addition to) ToolResultBlock content.
+            parent_id = sdk_message.parent_tool_use_id
+            if parent_id and parent_id not in resolved_in_blocks:
+                tool_info = self.current_tool_calls.get(parent_id, {})
+                tool_name = tool_info.get("name", "unknown")
+
+                # Try stashed output first (from PostToolUse hook),
+                # then tool_use_result dict, then string content.
+                output = pop_pending_tool_output(tool_name)
+                if not output:
+                    tur = sdk_message.tool_use_result
+                    if tur is not None:
+                        output = _extract_tool_use_result(tur)
+                if not output and isinstance(content, str) and content.strip():
+                    output = content.strip()
+
+                if output:
+                    responses.append(
+                        StreamToolOutputAvailable(
+                            toolCallId=parent_id,
+                            toolName=tool_name,
+                            output=output,
+                            success=True,
+                        )
+                    )
+                    resolved_in_blocks.add(parent_id)
+
+            self.resolved_tool_calls.update(resolved_in_blocks)
 
             # Close the current step after tool results — the next
             # AssistantMessage will open a new step for the continuation.
@@ -140,6 +178,7 @@ class SDKResponseAdapter:
                 self.step_open = False
 
         elif isinstance(sdk_message, ResultMessage):
+            self._flush_unresolved_tool_calls(responses)
             self._end_text_if_open(responses)
             # Close the step before finishing.
             if self.step_open:
@@ -149,7 +188,7 @@ class SDKResponseAdapter:
             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"
+                error_msg = sdk_message.result or "Unknown error"
                 responses.append(
                     StreamError(errorText=str(error_msg), code="sdk_error")
                 )
@@ -180,6 +219,59 @@ class SDKResponseAdapter:
             responses.append(StreamTextEnd(id=self.text_block_id))
             self.has_ended_text = True
 
+    def _flush_unresolved_tool_calls(self, responses: list[StreamBaseResponse]) -> None:
+        """Emit outputs for tool calls that didn't receive a UserMessage result.
+
+        SDK built-in tools (WebSearch, Read, etc.) may be executed by the CLI
+        internally without surfacing a separate ``UserMessage`` with
+        ``ToolResultBlock`` content.  The ``PostToolUse`` hook stashes their
+        output, which we pop and emit here before the next ``AssistantMessage``
+        starts.
+        """
+        flushed = False
+        for tool_id, tool_info in self.current_tool_calls.items():
+            if tool_id in self.resolved_tool_calls:
+                continue
+            tool_name = tool_info.get("name", "unknown")
+            output = pop_pending_tool_output(tool_name)
+            if output is not None:
+                responses.append(
+                    StreamToolOutputAvailable(
+                        toolCallId=tool_id,
+                        toolName=tool_name,
+                        output=output,
+                        success=True,
+                    )
+                )
+                self.resolved_tool_calls.add(tool_id)
+                flushed = True
+                logger.debug(
+                    f"Flushed pending output for built-in tool {tool_name} "
+                    f"(call {tool_id})"
+                )
+            else:
+                # No output available — emit an empty output so the frontend
+                # transitions the tool from input-available to output-available
+                # (stops the spinner).
+                responses.append(
+                    StreamToolOutputAvailable(
+                        toolCallId=tool_id,
+                        toolName=tool_name,
+                        output="",
+                        success=True,
+                    )
+                )
+                self.resolved_tool_calls.add(tool_id)
+                flushed = True
+                logger.debug(
+                    f"Flushed empty output for unresolved tool {tool_name} "
+                    f"(call {tool_id})"
+                )
+
+        if flushed and self.step_open:
+            responses.append(StreamFinishStep())
+            self.step_open = False
+
 
 def _extract_tool_output(content: str | list[dict[str, str]] | None) -> str:
     """Extract a string output from a ToolResultBlock's content field."""
@@ -199,3 +291,30 @@ def _extract_tool_output(content: str | list[dict[str, str]] | None) -> str:
         return json.dumps(content)
     except (TypeError, ValueError):
         return str(content)
+
+
+def _extract_tool_use_result(result: object) -> str:
+    """Extract a string from a UserMessage's ``tool_use_result`` dict.
+
+    SDK built-in tools may store their result in ``tool_use_result``
+    instead of (or in addition to) ``ToolResultBlock`` content blocks.
+    """
+    if isinstance(result, str):
+        return result
+    if isinstance(result, dict):
+        # Try common result keys
+        for key in ("content", "text", "output", "stdout", "result"):
+            val = result.get(key)
+            if isinstance(val, str) and val:
+                return val
+        # Fall back to JSON serialization of the whole dict
+        try:
+            return json.dumps(result)
+        except (TypeError, ValueError):
+            return str(result)
+    if result is None:
+        return ""
+    try:
+        return json.dumps(result)
+    except (TypeError, ValueError):
+        return str(result)

autogpt_platform/backend/backend/copilot/sdk/security_hooks.py

@@ -16,6 +16,7 @@ from .tool_adapter import (
     DANGEROUS_PATTERNS,
     MCP_TOOL_PREFIX,
     WORKSPACE_SCOPED_TOOLS,
+    stash_pending_tool_output,
 )
 
 logger = logging.getLogger(__name__)
@@ -224,10 +225,25 @@ def create_security_hooks(
             tool_use_id: str | None,
             context: HookContext,
         ) -> SyncHookJSONOutput:
-            """Log successful tool executions for observability."""
+            """Log successful tool executions and stash SDK built-in tool outputs.
+
+            MCP tools stash their output in ``_execute_tool_sync`` before the
+            SDK can truncate it.  SDK built-in tools (WebSearch, Read, etc.)
+            are executed by the CLI internally — this hook captures their
+            output so the response adapter can forward it to the frontend.
+            """
             _ = context
             tool_name = cast(str, input_data.get("tool_name", ""))
             logger.debug(f"[SDK] Tool success: {tool_name}, tool_use_id={tool_use_id}")
+
+            # Stash output for SDK built-in tools so the response adapter can
+            # emit StreamToolOutputAvailable even when the CLI doesn't surface
+            # a separate UserMessage with ToolResultBlock content.
+            if not tool_name.startswith(MCP_TOOL_PREFIX):
+                tool_response = input_data.get("tool_response")
+                if tool_response is not None:
+                    stash_pending_tool_output(tool_name, tool_response)
+
             return cast(SyncHookJSONOutput, {})
 
         async def post_tool_failure_hook(

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

@@ -47,6 +47,7 @@ from .tool_adapter import (
     set_execution_context,
 )
 from .transcript import (
+    cleanup_cli_project_dir,
     download_transcript,
     read_transcript_file,
     upload_transcript,
@@ -86,9 +87,12 @@ _SDK_TOOL_SUPPLEMENT = """
   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).
+- **IMPORTANT — File persistence**: Your working directory is **ephemeral** —
+  files are lost between turns. When you create or modify important files
+  (code, configs, outputs), you MUST save them using `write_workspace_file`
+  so they persist. Use `read_workspace_file` and `list_workspace_files` to
+  access files saved in previous turns. If a "Files from previous turns"
+  section is present above, those files are available via `read_workspace_file`.
 - 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.
@@ -268,48 +272,28 @@ def _make_sdk_cwd(session_id: str) -> str:
 
 
 def _cleanup_sdk_tool_results(cwd: str) -> None:
-    """Remove SDK tool-result files for a specific session working directory.
+    """Remove SDK session artifacts for a specific 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.
+    Cleans up:
+    - ``~/.claude/projects/<encoded-cwd>/`` — CLI session transcripts and
+      tool-result files.  Each SDK turn uses a unique cwd, so this directory
+      is safe to remove entirely.
+    - ``/tmp/copilot-<session>/`` — the ephemeral working directory.
 
-    Security: cwd MUST be created by _make_sdk_cwd() which sanitizes session_id.
+    Security: *cwd* MUST be created by ``_make_sdk_cwd()`` which sanitizes
+    the session_id.
     """
     import shutil
 
-    # 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 path outside workspace: {cwd}")
         return
 
-    # SDK encodes the cwd path by replacing '/' with '-'
-    encoded_cwd = normalized.replace("/", "-")
+    # Clean the CLI's project directory (transcripts + tool-results).
+    cleanup_cli_project_dir(cwd)
 
-    # 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
+    # Clean up the temp cwd directory itself.
     try:
         shutil.rmtree(normalized, ignore_errors=True)
     except OSError:
@@ -519,6 +503,7 @@ async def stream_chat_completion_sdk(
             def _on_stop(transcript_path: str, sdk_session_id: str) -> None:
                 captured_transcript.path = transcript_path
                 captured_transcript.sdk_session_id = sdk_session_id
+                logger.debug(f"[SDK] Stop hook: path={transcript_path!r}")
 
             security_hooks = create_security_hooks(
                 user_id,
@@ -530,18 +515,20 @@ async def stream_chat_completion_sdk(
             # --- Resume strategy: download transcript from bucket ---
             resume_file: str | None = None
             use_resume = False
+            transcript_msg_count = 0  # watermark: session.messages length at upload
 
             if config.claude_agent_use_resume and user_id and len(session.messages) > 1:
-                transcript_content = await download_transcript(user_id, session_id)
-                if transcript_content and validate_transcript(transcript_content):
+                dl = await download_transcript(user_id, session_id)
+                if dl and validate_transcript(dl.content):
                     resume_file = write_transcript_to_tempfile(
-                        transcript_content, session_id, sdk_cwd
+                        dl.content, session_id, sdk_cwd
                     )
                     if resume_file:
                         use_resume = True
-                        logger.info(
-                            f"[SDK] Using --resume with transcript "
-                            f"({len(transcript_content)} bytes)"
+                        transcript_msg_count = dl.message_count
+                        logger.debug(
+                            f"[SDK] Using --resume ({len(dl.content)}B, "
+                            f"msg_count={transcript_msg_count})"
                         )
 
             sdk_options_kwargs: dict[str, Any] = {
@@ -582,11 +569,35 @@ async def stream_chat_completion_sdk(
                 # Build query: with --resume the CLI already has full
                 # context, so we only send the new message.  Without
                 # resume, compress history into a context prefix.
+                #
+                # Hybrid mode: if the transcript is stale (upload missed
+                # some turns), compress only the gap and prepend it so
+                # the agent has transcript context + missed turns.
                 query_message = current_message
-                if not use_resume and len(session.messages) > 1:
+                current_msg_count = len(session.messages)
+
+                if use_resume and transcript_msg_count >= 0:
+                    # Transcript covers messages[0..M-1].  Current session
+                    # has N messages (last one is the new user msg).
+                    # Gap = messages[M .. N-2] (everything between upload
+                    # and the current turn).
+                    if transcript_msg_count < current_msg_count - 1:
+                        gap = session.messages[transcript_msg_count:-1]
+                        gap_context = _format_conversation_context(gap)
+                        if gap_context:
+                            logger.info(
+                                f"[SDK] Transcript stale: covers {transcript_msg_count} "
+                                f"of {current_msg_count} messages, compressing "
+                                f"{len(gap)} missed messages"
+                            )
+                            query_message = (
+                                f"{gap_context}\n\n"
+                                f"Now, the user says:\n{current_message}"
+                            )
+                elif not use_resume and current_msg_count > 1:
                     logger.warning(
                         f"[SDK] Using compression fallback for session "
-                        f"{session_id} ({len(session.messages)} messages) — "
+                        f"{session_id} ({current_msg_count} messages) — "
                         f"no transcript available for --resume"
                     )
                     compressed = await _compress_conversation_history(session)
@@ -597,10 +608,10 @@ async def stream_chat_completion_sdk(
                             f"Now, the user says:\n{current_message}"
                         )
 
-                logger.info(
-                    f"[SDK] Sending query ({len(session.messages)} msgs in session)"
+                logger.debug(
+                    f"[SDK] Sending query ({len(session.messages)} msgs, "
+                    f"resume={use_resume})"
                 )
-                logger.debug(f"[SDK] Query preview: {current_message[:80]!r}")
                 await client.query(query_message, session_id=session_id)
 
                 assistant_response = ChatMessage(role="assistant", content="")
@@ -681,29 +692,33 @@ async def stream_chat_completion_sdk(
                 ) and not has_appended_assistant:
                     session.messages.append(assistant_response)
 
-                # --- Capture transcript while CLI is still alive ---
-                # Must happen INSIDE async with: close() sends SIGTERM
-                # which kills the CLI before it can flush the JSONL.
-                if (
-                    config.claude_agent_use_resume
-                    and user_id
-                    and captured_transcript.available
-                ):
-                    # Give CLI time to flush JSONL writes before we read
-                    await asyncio.sleep(0.5)
+            # --- Upload transcript for next-turn --resume ---
+            # After async with the SDK task group has exited, so the Stop
+            # hook has already fired and the CLI has been SIGTERMed.  The
+            # CLI uses appendFileSync, so all writes are safely on disk.
+            if config.claude_agent_use_resume and user_id:
+                # With --resume the CLI appends to the resume file (most
+                # complete).  Otherwise use the Stop hook path.
+                if use_resume and resume_file:
+                    raw_transcript = read_transcript_file(resume_file)
+                elif captured_transcript.path:
                     raw_transcript = read_transcript_file(captured_transcript.path)
-                    if raw_transcript:
-                        try:
-                            async with asyncio.timeout(30):
-                                await _upload_transcript_bg(
-                                    user_id, session_id, raw_transcript
-                                )
-                        except asyncio.TimeoutError:
-                            logger.warning(
-                                f"[SDK] Transcript upload timed out for {session_id}"
-                            )
-                    else:
-                        logger.debug("[SDK] Stop hook fired but transcript not usable")
+                else:
+                    raw_transcript = None
+
+                if raw_transcript:
+                    # Shield the upload from generator cancellation so a
+                    # client disconnect / page refresh doesn't lose the
+                    # transcript.  The upload must finish even if the SSE
+                    # connection is torn down.
+                    await asyncio.shield(
+                        _try_upload_transcript(
+                            user_id,
+                            session_id,
+                            raw_transcript,
+                            message_count=len(session.messages),
+                        )
+                    )
 
         except ImportError:
             raise RuntimeError(
@@ -712,7 +727,7 @@ async def stream_chat_completion_sdk(
                 "to use the OpenAI-compatible fallback."
             )
 
-        await upsert_chat_session(session)
+        await asyncio.shield(upsert_chat_session(session))
         logger.debug(
             f"[SDK] Session {session_id} saved with {len(session.messages)} messages"
         )
@@ -722,7 +737,7 @@ async def stream_chat_completion_sdk(
     except Exception as e:
         logger.error(f"[SDK] Error: {e}", exc_info=True)
         try:
-            await upsert_chat_session(session)
+            await asyncio.shield(upsert_chat_session(session))
         except Exception as save_err:
             logger.error(f"[SDK] Failed to save session on error: {save_err}")
         yield StreamError(
@@ -735,14 +750,31 @@ async def stream_chat_completion_sdk(
             _cleanup_sdk_tool_results(sdk_cwd)
 
 
-async def _upload_transcript_bg(
-    user_id: str, session_id: str, raw_content: str
-) -> None:
-    """Background task to strip progress entries and upload transcript."""
+async def _try_upload_transcript(
+    user_id: str,
+    session_id: str,
+    raw_content: str,
+    message_count: int = 0,
+) -> bool:
+    """Strip progress entries and upload transcript (with timeout).
+
+    Returns True if the upload completed without error.
+    """
     try:
-        await upload_transcript(user_id, session_id, raw_content)
+        async with asyncio.timeout(30):
+            await upload_transcript(
+                user_id, session_id, raw_content, message_count=message_count
+            )
+        return True
+    except asyncio.TimeoutError:
+        logger.warning(f"[SDK] Transcript upload timed out for {session_id}")
+        return False
     except Exception as e:
-        logger.error(f"[SDK] Failed to upload transcript for {session_id}: {e}")
+        logger.error(
+            f"[SDK] Failed to upload transcript for {session_id}: {e}",
+            exc_info=True,
+        )
+        return False
 
 
 async def _update_title_async(

autogpt_platform/backend/backend/copilot/sdk/tool_adapter.py

@@ -41,7 +41,7 @@ _current_session: ContextVar[ChatSession | None] = ContextVar(
 # 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: ContextVar[dict[str, list[str]]] = ContextVar(
     "pending_tool_outputs", default=None  # type: ignore[arg-type]
 )
 
@@ -88,19 +88,52 @@ def get_execution_context() -> tuple[str | None, ChatSession | None]:
 
 
 def pop_pending_tool_output(tool_name: str) -> str | None:
-    """Pop and return the stashed full output for *tool_name*.
+    """Pop and return the oldest stashed 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.
 
+    Uses a FIFO queue per tool name so duplicate calls to the same tool
+    in one turn each get their own output.
+
     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)
+    queue = pending.get(tool_name)
+    if not queue:
+        pending.pop(tool_name, None)
+        return None
+    value = queue.pop(0)
+    if not queue:
+        del pending[tool_name]
+    return value
+
+
+def stash_pending_tool_output(tool_name: str, output: Any) -> None:
+    """Stash tool output for later retrieval by the response adapter.
+
+    Used by the PostToolUse hook to capture SDK built-in tool outputs
+    (WebSearch, Read, etc.) that aren't available through the MCP stash
+    mechanism in ``_execute_tool_sync``.
+
+    Appends to a FIFO queue per tool name so multiple calls to the same
+    tool in one turn are all preserved.
+    """
+    pending = _pending_tool_outputs.get(None)
+    if pending is None:
+        return
+    if isinstance(output, str):
+        text = output
+    else:
+        try:
+            text = json.dumps(output)
+        except (TypeError, ValueError):
+            text = str(output)
+    pending.setdefault(tool_name, []).append(text)
 
 
 async def _execute_tool_sync(
@@ -125,14 +158,63 @@ async def _execute_tool_sync(
     # 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
+        pending.setdefault(base_tool.name, []).append(text)
+
+    content_blocks: list[dict[str, str]] = [{"type": "text", "text": text}]
+
+    # If the tool result contains inline image data, add an MCP image block
+    # so Claude can "see" the image (e.g. read_workspace_file on a small PNG).
+    image_block = _extract_image_block(text)
+    if image_block:
+        content_blocks.append(image_block)
 
     return {
-        "content": [{"type": "text", "text": text}],
+        "content": content_blocks,
         "isError": not result.success,
     }
 
 
+# MIME types that Claude can process as image content blocks.
+_SUPPORTED_IMAGE_TYPES = frozenset(
+    {"image/png", "image/jpeg", "image/gif", "image/webp"}
+)
+
+
+def _extract_image_block(text: str) -> dict[str, str] | None:
+    """Extract an MCP image content block from a tool result JSON string.
+
+    Detects workspace file responses with ``content_base64`` and an image
+    MIME type, returning an MCP-format image block that allows Claude to
+    "see" the image.  Returns ``None`` if the result is not an inline image.
+    """
+    try:
+        data = json.loads(text)
+    except (json.JSONDecodeError, TypeError):
+        return None
+
+    if not isinstance(data, dict):
+        return None
+
+    mime_type = data.get("mime_type", "")
+    base64_content = data.get("content_base64", "")
+
+    # Only inline small images — large ones would exceed Claude's limits.
+    # 32 KB raw ≈ ~43 KB base64.
+    _MAX_IMAGE_BASE64_BYTES = 43_000
+    if (
+        mime_type in _SUPPORTED_IMAGE_TYPES
+        and base64_content
+        and len(base64_content) <= _MAX_IMAGE_BASE64_BYTES
+    ):
+        return {
+            "type": "image",
+            "data": base64_content,
+            "mimeType": mime_type,
+        }
+
+    return None
+
+
 def _mcp_error(message: str) -> dict[str, Any]:
     return {
         "content": [
@@ -311,14 +393,29 @@ def create_copilot_mcp_server():
 # which provides kernel-level network isolation via unshare --net.
 # Task allows spawning sub-agents (rate-limited by security hooks).
 # WebSearch uses Brave Search via Anthropic's API — safe, no SSRF risk.
-_SDK_BUILTIN_TOOLS = ["Read", "Write", "Edit", "Glob", "Grep", "Task", "WebSearch"]
+# TodoWrite manages the task checklist shown in the UI — no security concern.
+_SDK_BUILTIN_TOOLS = [
+    "Read",
+    "Write",
+    "Edit",
+    "Glob",
+    "Grep",
+    "Task",
+    "WebSearch",
+    "TodoWrite",
+]
 
 # SDK built-in tools that must be explicitly blocked.
 # Bash: dangerous — agent uses mcp__copilot__bash_exec with kernel-level
 #   network isolation (unshare --net) instead.
 # WebFetch: SSRF risk — can reach internal network (localhost, 10.x, etc.).
 #   Agent uses the SSRF-protected mcp__copilot__web_fetch tool instead.
-SDK_DISALLOWED_TOOLS = ["Bash", "WebFetch"]
+# AskUserQuestion: interactive CLI tool — no terminal in copilot context.
+SDK_DISALLOWED_TOOLS = [
+    "Bash",
+    "WebFetch",
+    "AskUserQuestion",
+]
 
 # Tools that are blocked entirely in security hooks (defence-in-depth).
 # Includes SDK_DISALLOWED_TOOLS plus common aliases/synonyms.

autogpt_platform/backend/backend/copilot/sdk/transcript.py

@@ -14,6 +14,8 @@ import json
 import logging
 import os
 import re
+import time
+from dataclasses import dataclass
 
 logger = logging.getLogger(__name__)
 
@@ -31,6 +33,16 @@ STRIPPABLE_TYPES = frozenset(
     {"progress", "file-history-snapshot", "queue-operation", "summary", "pr-link"}
 )
 
+
+@dataclass
+class TranscriptDownload:
+    """Result of downloading a transcript with its metadata."""
+
+    content: str
+    message_count: int = 0  # session.messages length when uploaded
+    uploaded_at: float = 0.0  # epoch timestamp of upload
+
+
 # Workspace storage constants — deterministic path from session_id.
 TRANSCRIPT_STORAGE_PREFIX = "chat-transcripts"
 
@@ -119,23 +131,19 @@ def read_transcript_file(transcript_path: str) -> str | None:
             content = f.read()
 
         if not content.strip():
-            logger.debug(f"[Transcript] Empty file: {transcript_path}")
             return None
 
         lines = content.strip().split("\n")
         if len(lines) < 3:
             # Raw files with ≤2 lines are metadata-only
             # (queue-operation + file-history-snapshot, no conversation).
-            logger.debug(
-                f"[Transcript] Too few lines ({len(lines)}): {transcript_path}"
-            )
             return None
 
         # Quick structural validation — parse first and last lines.
         json.loads(lines[0])
         json.loads(lines[-1])
 
-        logger.info(
+        logger.debug(
             f"[Transcript] Read {len(lines)} lines, "
             f"{len(content)} bytes from {transcript_path}"
         )
@@ -160,6 +168,41 @@ def _sanitize_id(raw_id: str, max_len: int = 36) -> str:
 _SAFE_CWD_PREFIX = os.path.realpath("/tmp/copilot-")
 
 
+def _encode_cwd_for_cli(cwd: str) -> str:
+    """Encode a working directory path the same way the Claude CLI does.
+
+    The CLI replaces all non-alphanumeric characters with ``-``.
+    """
+    return re.sub(r"[^a-zA-Z0-9]", "-", os.path.realpath(cwd))
+
+
+def cleanup_cli_project_dir(sdk_cwd: str) -> None:
+    """Remove the CLI's project directory for a specific working directory.
+
+    The CLI stores session data under ``~/.claude/projects/<encoded_cwd>/``.
+    Each SDK turn uses a unique ``sdk_cwd``, so the project directory is
+    safe to remove entirely after the transcript has been uploaded.
+    """
+    import shutil
+
+    cwd_encoded = _encode_cwd_for_cli(sdk_cwd)
+    config_dir = os.environ.get("CLAUDE_CONFIG_DIR") or os.path.expanduser("~/.claude")
+    projects_base = os.path.realpath(os.path.join(config_dir, "projects"))
+    project_dir = os.path.realpath(os.path.join(projects_base, cwd_encoded))
+
+    if not project_dir.startswith(projects_base + os.sep):
+        logger.warning(
+            f"[Transcript] Cleanup path escaped projects base: {project_dir}"
+        )
+        return
+
+    if os.path.isdir(project_dir):
+        shutil.rmtree(project_dir, ignore_errors=True)
+        logger.debug(f"[Transcript] Cleaned up CLI project dir: {project_dir}")
+    else:
+        logger.debug(f"[Transcript] Project dir not found: {project_dir}")
+
+
 def write_transcript_to_tempfile(
     transcript_content: str,
     session_id: str,
@@ -191,7 +234,7 @@ def write_transcript_to_tempfile(
         with open(jsonl_path, "w") as f:
             f.write(transcript_content)
 
-        logger.info(f"[Transcript] Wrote resume file: {jsonl_path}")
+        logger.debug(f"[Transcript] Wrote resume file: {jsonl_path}")
         return jsonl_path
 
     except OSError as e:
@@ -248,6 +291,15 @@ def _storage_path_parts(user_id: str, session_id: str) -> tuple[str, str, str]:
     )
 
 
+def _meta_storage_path_parts(user_id: str, session_id: str) -> tuple[str, str, str]:
+    """Return (workspace_id, file_id, filename) for a session's transcript metadata."""
+    return (
+        TRANSCRIPT_STORAGE_PREFIX,
+        _sanitize_id(user_id),
+        f"{_sanitize_id(session_id)}.meta.json",
+    )
+
+
 def _build_storage_path(user_id: str, session_id: str, backend: object) -> str:
     """Build the full storage path string that ``retrieve()`` expects.
 
@@ -268,21 +320,30 @@ def _build_storage_path(user_id: str, session_id: str, backend: object) -> str:
         return f"local://{wid}/{fid}/{fname}"
 
 
-async def upload_transcript(user_id: str, session_id: str, content: str) -> None:
+async def upload_transcript(
+    user_id: str,
+    session_id: str,
+    content: str,
+    message_count: int = 0,
+) -> None:
     """Strip progress entries and upload transcript to bucket storage.
 
     Safety: only overwrites when the new (stripped) transcript is larger than
     what is already stored.  Since JSONL is append-only, the latest transcript
     is always the longest.  This prevents a slow/stale background task from
     clobbering a newer upload from a concurrent turn.
+
+    Args:
+        message_count: ``len(session.messages)`` at upload time — used by
+            the next turn to detect staleness and compress only the gap.
     """
     from backend.util.workspace_storage import get_workspace_storage
 
     stripped = strip_progress_entries(content)
     if not validate_transcript(stripped):
         logger.warning(
-            f"[Transcript] Skipping upload — stripped content is not a valid "
-            f"transcript for session {session_id}"
+            f"[Transcript] Skipping upload — stripped content not valid "
+            f"for session {session_id}"
         )
         return
 
@@ -296,10 +357,9 @@ async def upload_transcript(user_id: str, session_id: str, content: str) -> None
     try:
         existing = await storage.retrieve(path)
         if len(existing) >= new_size:
-            logger.info(
-                f"[Transcript] Skipping upload — existing transcript "
-                f"({len(existing)}B) >= new ({new_size}B) for session "
-                f"{session_id}"
+            logger.debug(
+                f"[Transcript] Skipping upload — existing ({len(existing)}B) "
+                f">= new ({new_size}B) for session {session_id}"
             )
             return
     except (FileNotFoundError, Exception):
@@ -311,16 +371,38 @@ async def upload_transcript(user_id: str, session_id: str, content: str) -> None
         filename=fname,
         content=encoded,
     )
+
+    # Store metadata alongside the transcript so the next turn can detect
+    # staleness and only compress the gap instead of the full history.
+    # Wrapped in try/except so a metadata write failure doesn't orphan
+    # the already-uploaded transcript — the next turn will just fall back
+    # to full gap fill (msg_count=0).
+    try:
+        meta = {"message_count": message_count, "uploaded_at": time.time()}
+        mwid, mfid, mfname = _meta_storage_path_parts(user_id, session_id)
+        await storage.store(
+            workspace_id=mwid,
+            file_id=mfid,
+            filename=mfname,
+            content=json.dumps(meta).encode("utf-8"),
+        )
+    except Exception as e:
+        logger.warning(f"[Transcript] Failed to write metadata for {session_id}: {e}")
+
     logger.info(
-        f"[Transcript] Uploaded {new_size} bytes "
-        f"(stripped from {len(content)}) for session {session_id}"
+        f"[Transcript] Uploaded {new_size}B "
+        f"(stripped from {len(content)}B, msg_count={message_count}) "
+        f"for session {session_id}"
     )
 
 
-async def download_transcript(user_id: str, session_id: str) -> str | None:
-    """Download transcript from bucket storage.
+async def download_transcript(
+    user_id: str, session_id: str
+) -> TranscriptDownload | None:
+    """Download transcript and metadata from bucket storage.
 
-    Returns the JSONL content string, or ``None`` if not found.
+    Returns a ``TranscriptDownload`` with the JSONL content and the
+    ``message_count`` watermark from the upload, or ``None`` if not found.
     """
     from backend.util.workspace_storage import get_workspace_storage
 
@@ -330,10 +412,6 @@ async def download_transcript(user_id: str, session_id: str) -> str | None:
     try:
         data = await storage.retrieve(path)
         content = data.decode("utf-8")
-        logger.info(
-            f"[Transcript] Downloaded {len(content)} bytes for session {session_id}"
-        )
-        return content
     except FileNotFoundError:
         logger.debug(f"[Transcript] No transcript in storage for {session_id}")
         return None
@@ -341,6 +419,36 @@ async def download_transcript(user_id: str, session_id: str) -> str | None:
         logger.warning(f"[Transcript] Failed to download transcript: {e}")
         return None
 
+    # Try to load metadata (best-effort — old transcripts won't have it)
+    message_count = 0
+    uploaded_at = 0.0
+    try:
+        from backend.util.workspace_storage import GCSWorkspaceStorage
+
+        mwid, mfid, mfname = _meta_storage_path_parts(user_id, session_id)
+        if isinstance(storage, GCSWorkspaceStorage):
+            blob = f"workspaces/{mwid}/{mfid}/{mfname}"
+            meta_path = f"gcs://{storage.bucket_name}/{blob}"
+        else:
+            meta_path = f"local://{mwid}/{mfid}/{mfname}"
+
+        meta_data = await storage.retrieve(meta_path)
+        meta = json.loads(meta_data.decode("utf-8"))
+        message_count = meta.get("message_count", 0)
+        uploaded_at = meta.get("uploaded_at", 0.0)
+    except (FileNotFoundError, json.JSONDecodeError, Exception):
+        pass  # No metadata — treat as unknown (msg_count=0 → always fill gap)
+
+    logger.debug(
+        f"[Transcript] Downloaded {len(content)}B "
+        f"(msg_count={message_count}) for session {session_id}"
+    )
+    return TranscriptDownload(
+        content=content,
+        message_count=message_count,
+        uploaded_at=uploaded_at,
+    )
+
 
 async def delete_transcript(user_id: str, session_id: str) -> None:
     """Delete transcript from bucket storage (e.g. after resume failure)."""

autogpt_platform/backend/backend/copilot/service.py

@@ -387,7 +387,7 @@ async def stream_chat_completion(
     if user_id:
         log_meta["user_id"] = user_id
 
-    logger.info(
+    logger.debug(
         f"[TIMING] stream_chat_completion STARTED, session={session_id}, user={user_id}, "
         f"message_len={len(message) if message else 0}, is_user={is_user_message}",
         extra={
@@ -404,7 +404,7 @@ async def stream_chat_completion(
         fetch_start = time.monotonic()
         session = await get_chat_session(session_id, user_id)
         fetch_time = (time.monotonic() - fetch_start) * 1000
-        logger.info(
+        logger.debug(
             f"[TIMING] get_chat_session took {fetch_time:.1f}ms, "
             f"n_messages={len(session.messages) if session else 0}",
             extra={
@@ -416,7 +416,7 @@ async def stream_chat_completion(
             },
         )
     else:
-        logger.info(
+        logger.debug(
             f"[TIMING] Using provided session, messages={len(session.messages)}",
             extra={"json_fields": {**log_meta, "n_messages": len(session.messages)}},
         )
@@ -450,7 +450,7 @@ async def stream_chat_completion(
                 message_length=len(message),
             )
             posthog_time = (time.monotonic() - posthog_start) * 1000
-            logger.info(
+            logger.debug(
                 f"[TIMING] track_user_message took {posthog_time:.1f}ms",
                 extra={"json_fields": {**log_meta, "duration_ms": posthog_time}},
             )
@@ -458,7 +458,7 @@ async def stream_chat_completion(
     upsert_start = time.monotonic()
     session = await upsert_chat_session(session)
     upsert_time = (time.monotonic() - upsert_start) * 1000
-    logger.info(
+    logger.debug(
         f"[TIMING] upsert_chat_session took {upsert_time:.1f}ms",
         extra={"json_fields": {**log_meta, "duration_ms": upsert_time}},
     )
@@ -503,7 +503,7 @@ async def stream_chat_completion(
     prompt_start = time.monotonic()
     system_prompt, understanding = await _build_system_prompt(user_id)
     prompt_time = (time.monotonic() - prompt_start) * 1000
-    logger.info(
+    logger.debug(
         f"[TIMING] _build_system_prompt took {prompt_time:.1f}ms",
         extra={"json_fields": {**log_meta, "duration_ms": prompt_time}},
     )
@@ -537,7 +537,7 @@ async def stream_chat_completion(
 
     # Only yield message start for the initial call, not for continuations.
     setup_time = (time.monotonic() - completion_start) * 1000
-    logger.info(
+    logger.debug(
         f"[TIMING] Setup complete, yielding StreamStart at {setup_time:.1f}ms",
         extra={"json_fields": {**log_meta, "setup_time_ms": setup_time}},
     )
@@ -548,7 +548,7 @@ async def stream_chat_completion(
     yield StreamStartStep()
 
     try:
-        logger.info(
+        logger.debug(
             "[TIMING] Calling _stream_chat_chunks",
             extra={"json_fields": log_meta},
         )
@@ -988,7 +988,7 @@ async def _stream_chat_chunks(
     if session.user_id:
         log_meta["user_id"] = session.user_id
 
-    logger.info(
+    logger.debug(
         f"[TIMING] _stream_chat_chunks STARTED, session={session.session_id}, "
         f"user={session.user_id}, n_messages={len(session.messages)}",
         extra={"json_fields": {**log_meta, "n_messages": len(session.messages)}},
@@ -1011,7 +1011,7 @@ async def _stream_chat_chunks(
         base_url=config.base_url,
     )
     context_time = (time_module.perf_counter() - context_start) * 1000
-    logger.info(
+    logger.debug(
         f"[TIMING] _manage_context_window took {context_time:.1f}ms",
         extra={"json_fields": {**log_meta, "duration_ms": context_time}},
     )
@@ -1053,7 +1053,7 @@ async def _stream_chat_chunks(
                 retry_info = (
                     f" (retry {retry_count}/{MAX_RETRIES})" if retry_count > 0 else ""
                 )
-                logger.info(
+                logger.debug(
                     f"[TIMING] Creating OpenAI stream at {elapsed:.1f}ms{retry_info}",
                     extra={
                         "json_fields": {
@@ -1093,7 +1093,7 @@ async def _stream_chat_chunks(
                     extra_body=extra_body,
                 )
                 api_init_time = (time_module.perf_counter() - api_call_start) * 1000
-                logger.info(
+                logger.debug(
                     f"[TIMING] OpenAI stream object returned in {api_init_time:.1f}ms",
                     extra={"json_fields": {**log_meta, "duration_ms": api_init_time}},
                 )
@@ -1142,7 +1142,7 @@ async def _stream_chat_chunks(
                                 ttfc = (
                                     time_module.perf_counter() - api_call_start
                                 ) * 1000
-                                logger.info(
+                                logger.debug(
                                     f"[TIMING] FIRST CONTENT CHUNK at {ttfc:.1f}ms "
                                     f"(since API call), n_chunks={chunk_count}",
                                     extra={
@@ -1210,7 +1210,7 @@ async def _stream_chat_chunks(
                                     )
                                     emitted_start_for_idx.add(idx)
                 stream_duration = time_module.perf_counter() - api_call_start
-                logger.info(
+                logger.debug(
                     f"[TIMING] OpenAI stream COMPLETE, finish_reason={finish_reason}, "
                     f"duration={stream_duration:.2f}s, "
                     f"n_chunks={chunk_count}, n_tool_calls={len(tool_calls)}",
@@ -1244,7 +1244,7 @@ async def _stream_chat_chunks(
                         raise
 
                 total_time = (time_module.perf_counter() - stream_chunks_start) * 1000
-                logger.info(
+                logger.debug(
                     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}},
@@ -1494,8 +1494,8 @@ async def _yield_tool_call(
             # Mark stream registry task as failed if it was created
             try:
                 await stream_registry.mark_task_completed(task_id, status="failed")
-            except Exception:
-                pass
+            except Exception as mark_err:
+                logger.warning(f"Failed to mark task {task_id} as failed: {mark_err}")
             logger.error(
                 f"Failed to setup long-running tool {tool_name}: {e}", exc_info=True
             )

autogpt_platform/backend/backend/copilot/service_test.py

@@ -143,7 +143,7 @@ async def test_sdk_resume_multi_turn(setup_test_user, test_user_id):
         "Transcript was not uploaded to bucket after turn 1 — "
         "Stop hook may not have fired or transcript was too small"
     )
-    logger.info(f"Turn 1 transcript uploaded: {len(transcript)} bytes")
+    logger.info(f"Turn 1 transcript uploaded: {len(transcript.content)} bytes")
 
     # Reload session for turn 2
     session = await get_chat_session(session.session_id, test_user_id)

autogpt_platform/backend/backend/copilot/stream_registry.py

@@ -117,7 +117,7 @@ async def create_task(
     if user_id:
         log_meta["user_id"] = user_id
 
-    logger.info(
+    logger.debug(
         f"[TIMING] create_task STARTED, task={task_id}, session={session_id}, user={user_id}",
         extra={"json_fields": log_meta},
     )
@@ -135,7 +135,7 @@ async def create_task(
     redis_start = time.perf_counter()
     redis = await get_redis_async()
     redis_time = (time.perf_counter() - redis_start) * 1000
-    logger.info(
+    logger.debug(
         f"[TIMING] get_redis_async took {redis_time:.1f}ms",
         extra={"json_fields": {**log_meta, "duration_ms": redis_time}},
     )
@@ -158,7 +158,7 @@ async def create_task(
         },
     )
     hset_time = (time.perf_counter() - hset_start) * 1000
-    logger.info(
+    logger.debug(
         f"[TIMING] redis.hset took {hset_time:.1f}ms",
         extra={"json_fields": {**log_meta, "duration_ms": hset_time}},
     )
@@ -169,7 +169,7 @@ async def create_task(
     await redis.set(op_key, task_id, ex=config.stream_ttl)
 
     total_time = (time.perf_counter() - start_time) * 1000
-    logger.info(
+    logger.debug(
         f"[TIMING] create_task COMPLETED in {total_time:.1f}ms; task={task_id}, session={session_id}",
         extra={"json_fields": {**log_meta, "total_time_ms": total_time}},
     )
@@ -230,7 +230,7 @@ async def publish_chunk(
             in ("StreamStart", "StreamFinish", "StreamTextStart", "StreamTextEnd")
             or total_time > 50
         ):
-            logger.info(
+            logger.debug(
                 f"[TIMING] publish_chunk {chunk_type} in {total_time:.1f}ms (xadd={xadd_time:.1f}ms)",
                 extra={
                     "json_fields": {
@@ -279,7 +279,7 @@ async def subscribe_to_task(
     if user_id:
         log_meta["user_id"] = user_id
 
-    logger.info(
+    logger.debug(
         f"[TIMING] subscribe_to_task STARTED, task={task_id}, user={user_id}, last_msg={last_message_id}",
         extra={"json_fields": {**log_meta, "last_message_id": last_message_id}},
     )
@@ -289,14 +289,14 @@ async def subscribe_to_task(
     meta_key = _get_task_meta_key(task_id)
     meta: dict[Any, Any] = await redis.hgetall(meta_key)  # type: ignore[misc]
     hgetall_time = (time.perf_counter() - redis_start) * 1000
-    logger.info(
+    logger.debug(
         f"[TIMING] Redis hgetall took {hgetall_time:.1f}ms",
         extra={"json_fields": {**log_meta, "duration_ms": hgetall_time}},
     )
 
     if not meta:
         elapsed = (time.perf_counter() - start_time) * 1000
-        logger.info(
+        logger.debug(
             f"[TIMING] Task not found in Redis after {elapsed:.1f}ms",
             extra={
                 "json_fields": {
@@ -335,7 +335,7 @@ async def subscribe_to_task(
     xread_start = time.perf_counter()
     messages = await redis.xread({stream_key: last_message_id}, block=0, count=1000)
     xread_time = (time.perf_counter() - xread_start) * 1000
-    logger.info(
+    logger.debug(
         f"[TIMING] Redis xread (replay) took {xread_time:.1f}ms, status={task_status}",
         extra={
             "json_fields": {
@@ -363,7 +363,7 @@ async def subscribe_to_task(
                     except Exception as e:
                         logger.warning(f"Failed to replay message: {e}")
 
-    logger.info(
+    logger.debug(
         f"[TIMING] Replayed {replayed_count} messages, last_id={replay_last_id}",
         extra={
             "json_fields": {
@@ -376,7 +376,7 @@ async def subscribe_to_task(
 
     # Step 2: If task is still running, start stream listener for live updates
     if task_status == "running":
-        logger.info(
+        logger.debug(
             "[TIMING] Task still running, starting _stream_listener",
             extra={"json_fields": {**log_meta, "task_status": task_status}},
         )
@@ -387,14 +387,14 @@ async def subscribe_to_task(
         _listener_tasks[id(subscriber_queue)] = (task_id, listener_task)
     else:
         # Task is completed/failed - add finish marker
-        logger.info(
+        logger.debug(
             f"[TIMING] Task already {task_status}, adding StreamFinish",
             extra={"json_fields": {**log_meta, "task_status": task_status}},
         )
         await subscriber_queue.put(StreamFinish())
 
     total_time = (time.perf_counter() - start_time) * 1000
-    logger.info(
+    logger.debug(
         f"[TIMING] subscribe_to_task COMPLETED in {total_time:.1f}ms; task={task_id}, "
         f"n_messages_replayed={replayed_count}",
         extra={
@@ -433,7 +433,7 @@ async def _stream_listener(
     if log_meta is None:
         log_meta = {"component": "StreamRegistry", "task_id": task_id}
 
-    logger.info(
+    logger.debug(
         f"[TIMING] _stream_listener STARTED, task={task_id}, last_id={last_replayed_id}",
         extra={"json_fields": {**log_meta, "last_replayed_id": last_replayed_id}},
     )
@@ -462,7 +462,7 @@ async def _stream_listener(
 
             if messages:
                 msg_count = sum(len(msgs) for _, msgs in messages)
-                logger.info(
+                logger.debug(
                     f"[TIMING] xread #{xread_count} returned {msg_count} messages in {xread_time:.1f}ms",
                     extra={
                         "json_fields": {
@@ -475,7 +475,7 @@ async def _stream_listener(
                 )
             elif xread_time > 1000:
                 # Only log timeouts (30s blocking)
-                logger.info(
+                logger.debug(
                     f"[TIMING] xread #{xread_count} timeout after {xread_time:.1f}ms",
                     extra={
                         "json_fields": {
@@ -526,7 +526,7 @@ async def _stream_listener(
                                 if first_message_time is None:
                                     first_message_time = time.perf_counter()
                                     elapsed = (first_message_time - start_time) * 1000
-                                    logger.info(
+                                    logger.debug(
                                         f"[TIMING] FIRST live message at {elapsed:.1f}ms, type={type(chunk).__name__}",
                                         extra={
                                             "json_fields": {
@@ -568,7 +568,7 @@ async def _stream_listener(
                             # Stop listening on finish
                             if isinstance(chunk, StreamFinish):
                                 total_time = (time.perf_counter() - start_time) * 1000
-                                logger.info(
+                                logger.debug(
                                     f"[TIMING] StreamFinish received in {total_time/1000:.1f}s; delivered={messages_delivered}",
                                     extra={
                                         "json_fields": {
@@ -587,7 +587,7 @@ async def _stream_listener(
 
     except asyncio.CancelledError:
         elapsed = (time.perf_counter() - start_time) * 1000
-        logger.info(
+        logger.debug(
             f"[TIMING] _stream_listener CANCELLED after {elapsed:.1f}ms, delivered={messages_delivered}",
             extra={
                 "json_fields": {
@@ -619,7 +619,7 @@ async def _stream_listener(
     finally:
         # Clean up listener task mapping on exit
         total_time = (time.perf_counter() - start_time) * 1000
-        logger.info(
+        logger.debug(
             f"[TIMING] _stream_listener FINISHED in {total_time/1000:.1f}s; task={task_id}, "
             f"delivered={messages_delivered}, xread_count={xread_count}",
             extra={
@@ -829,10 +829,13 @@ async def get_active_task_for_session(
                             )
                             await mark_task_completed(task_id, "failed")
                             continue
-                    except (ValueError, TypeError):
-                        pass
+                    except (ValueError, TypeError) as exc:
+                        logger.warning(
+                            f"[TASK_LOOKUP] Failed to parse created_at "
+                            f"for task {task_id[:8]}...: {exc}"
+                        )
 
-                logger.info(
+                logger.debug(
                     f"[TASK_LOOKUP] Found running task {task_id[:8]}... for session {session_id[:8]}..."
                 )
 

autogpt_platform/backend/backend/copilot/tools/feature_requests.py

@@ -33,7 +33,6 @@ query SearchFeatureRequests($term: String!, $filter: IssueFilter, $first: Int) {
       id
       identifier
       title
-      description
     }
   }
 }
@@ -205,7 +204,6 @@ class SearchFeatureRequestsTool(BaseTool):
                     id=node["id"],
                     identifier=node["identifier"],
                     title=node["title"],
-                    description=node.get("description"),
                 )
                 for node in nodes
             ]
@@ -239,7 +237,11 @@ class CreateFeatureRequestTool(BaseTool):
             "Create a new feature request or add a customer need to an existing one. "
             "Always search first with search_feature_requests to avoid duplicates. "
             "If a matching request exists, pass its ID as existing_issue_id to add "
-            "the user's need to it instead of creating a duplicate."
+            "the user's need to it instead of creating a duplicate. "
+            "IMPORTANT: Never include personally identifiable information (PII) in "
+            "the title or description — no names, emails, phone numbers, company "
+            "names, or other identifying details. Write titles and descriptions in "
+            "generic, feature-focused language."
         )
 
     @property
@@ -249,11 +251,20 @@ class CreateFeatureRequestTool(BaseTool):
             "properties": {
                 "title": {
                     "type": "string",
-                    "description": "Title for the feature request.",
+                    "description": (
+                        "Title for the feature request. Must be generic and "
+                        "feature-focused — do not include any user names, emails, "
+                        "company names, or other PII."
+                    ),
                 },
                 "description": {
                     "type": "string",
-                    "description": "Detailed description of what the user wants and why.",
+                    "description": (
+                        "Detailed description of what the user wants and why. "
+                        "Must not contain any personally identifiable information "
+                        "(PII) — describe the feature need generically without "
+                        "referencing specific users, companies, or contact details."
+                    ),
                 },
                 "existing_issue_id": {
                     "type": "string",

autogpt_platform/backend/backend/copilot/tools/feature_requests_test.py

@@ -117,13 +117,11 @@ class TestSearchFeatureRequestsTool:
                 "id": "id-1",
                 "identifier": "FR-1",
                 "title": "Dark mode",
-                "description": "Add dark mode support",
             },
             {
                 "id": "id-2",
                 "identifier": "FR-2",
                 "title": "Dark theme",
-                "description": None,
             },
         ]
         patcher, _ = _mock_linear_config(query_return=_search_response(nodes))

autogpt_platform/backend/backend/copilot/tools/models.py

@@ -486,7 +486,6 @@ class FeatureRequestInfo(BaseModel):
     id: str
     identifier: str
     title: str
-    description: str | None = None
 
 
 class FeatureRequestSearchResponse(ToolResponseBase):

autogpt_platform/backend/backend/copilot/tools/workspace_files.py

@@ -312,8 +312,18 @@ class ReadWorkspaceFileTool(BaseTool):
             is_small_file = file_info.size_bytes <= self.MAX_INLINE_SIZE_BYTES
             is_text_file = self._is_text_mime_type(file_info.mime_type)
 
-            # Return inline content for small text files (unless force_download_url)
-            if is_small_file and is_text_file and not force_download_url:
+            # Return inline content for small text/image files (unless force_download_url)
+            is_image_file = file_info.mime_type in {
+                "image/png",
+                "image/jpeg",
+                "image/gif",
+                "image/webp",
+            }
+            if (
+                is_small_file
+                and (is_text_file or is_image_file)
+                and not force_download_url
+            ):
                 content = await manager.read_file_by_id(target_file_id)
                 content_b64 = base64.b64encode(content).decode("utf-8")
 

autogpt_platform/backend/backend/util/service.py

@@ -599,6 +599,15 @@ def get_service_client(
                 if error_response and error_response.type in EXCEPTION_MAPPING:
                     exception_class = EXCEPTION_MAPPING[error_response.type]
                     args = error_response.args or [str(e)]
+
+                    # Prisma DataError subclasses expect a dict `data` arg,
+                    # but RPC serialization only preserves the string message
+                    # from exc.args.  Wrap it in the expected structure so
+                    # the constructor doesn't crash on `.get()`.
+                    if issubclass(exception_class, DataError):
+                        msg = str(args[0]) if args else str(e)
+                        raise exception_class({"user_facing_error": {"message": msg}})
+
                     raise exception_class(*args)
 
                 # Otherwise categorize by HTTP status code

autogpt_platform/backend/backend/util/service_test.py

@@ -6,6 +6,7 @@ from unittest.mock import Mock
 
 import httpx
 import pytest
+from prisma.errors import DataError, UniqueViolationError
 
 from backend.util.service import (
     AppService,
@@ -447,6 +448,39 @@ class TestHTTPErrorRetryBehavior:
 
         assert "Invalid parameter value" in str(exc_info.value)
 
+    def test_prisma_data_error_reconstructed_correctly(self):
+        """Test that DataError subclasses (e.g. UniqueViolationError) are
+        reconstructed without crashing.
+
+        Prisma's DataError.__init__ expects a dict `data` arg with
+        a 'user_facing_error' key.  RPC serialization only preserves the
+        string message via exc.args, so the client must wrap it in the
+        expected dict structure.
+        """
+        for exc_type in [DataError, UniqueViolationError]:
+            mock_response = Mock()
+            mock_response.status_code = 400
+            mock_response.json.return_value = {
+                "type": exc_type.__name__,
+                "args": ["Unique constraint failed on the fields: (`path`)"],
+            }
+            mock_response.raise_for_status.side_effect = httpx.HTTPStatusError(
+                "400 Bad Request", request=Mock(), response=mock_response
+            )
+
+            client = get_service_client(ServiceTestClient)
+
+            with pytest.raises(exc_type) as exc_info:
+                client._handle_call_method_response(  # type: ignore[attr-defined]
+                    response=mock_response, method_name="test_method"
+                )
+
+            # The exception should have the message preserved
+            assert "Unique constraint" in str(exc_info.value)
+            # And should have the expected data structure (not crash)
+            assert hasattr(exc_info.value, "data")
+            assert isinstance(exc_info.value.data, dict)
+
     def test_client_error_status_codes_coverage(self):
         """Test that various 4xx status codes are all wrapped as HTTPClientError."""
         client_error_codes = [400, 401, 403, 404, 405, 409, 422, 429]

autogpt_platform/backend/backend/util/workspace_storage.py

@@ -93,7 +93,14 @@ class WorkspaceStorageBackend(ABC):
 
 
 class GCSWorkspaceStorage(WorkspaceStorageBackend):
-    """Google Cloud Storage implementation for workspace storage."""
+    """Google Cloud Storage implementation for workspace storage.
+
+    Each instance owns a single ``aiohttp.ClientSession`` and GCS async
+    client.  Because ``ClientSession`` is bound to the event loop on which it
+    was created, callers that run on separate loops (e.g. copilot executor
+    worker threads) **must** obtain their own ``GCSWorkspaceStorage`` instance
+    via :func:`get_workspace_storage` which is event-loop-aware.
+    """
 
     def __init__(self, bucket_name: str):
         self.bucket_name = bucket_name
@@ -337,60 +344,73 @@ class LocalWorkspaceStorage(WorkspaceStorageBackend):
             raise ValueError(f"Invalid storage path format: {storage_path}")
 
 
-# Global storage backend instance
-_workspace_storage: Optional[WorkspaceStorageBackend] = None
+# ---------------------------------------------------------------------------
+# Storage instance management
+# ---------------------------------------------------------------------------
+# ``aiohttp.ClientSession`` is bound to the event loop where it is created.
+# The copilot executor runs each worker in its own thread with a dedicated
+# event loop, so a single global ``GCSWorkspaceStorage`` instance would break.
+#
+# For **local storage** a single shared instance is fine (no async I/O).
+# For **GCS storage** we keep one instance *per event loop* so every loop
+# gets its own ``ClientSession``.
+# ---------------------------------------------------------------------------
+
+_local_storage: Optional[LocalWorkspaceStorage] = None
+_gcs_storages: dict[int, GCSWorkspaceStorage] = {}
 _storage_lock = asyncio.Lock()
 
 
 async def get_workspace_storage() -> WorkspaceStorageBackend:
+    """Return a workspace storage backend for the **current** event loop.
+
+    * Local storage → single shared instance (no event-loop affinity).
+    * GCS storage   → one instance per event loop to avoid cross-loop
+      ``aiohttp`` errors.
     """
-    Get the workspace storage backend instance.
+    global _local_storage
 
-    Uses GCS if media_gcs_bucket_name is configured, otherwise uses local storage.
-    """
-    global _workspace_storage
+    config = Config()
 
-    if _workspace_storage is None:
-        async with _storage_lock:
-            if _workspace_storage is None:
-                config = Config()
+    # --- Local storage (shared) ---
+    if not config.media_gcs_bucket_name:
+        if _local_storage is None:
+            storage_dir = (
+                config.workspace_storage_dir if config.workspace_storage_dir else None
+            )
+            logger.info(f"Using local workspace storage: {storage_dir or 'default'}")
+            _local_storage = LocalWorkspaceStorage(storage_dir)
+        return _local_storage
 
-                if config.media_gcs_bucket_name:
-                    logger.info(
-                        f"Using GCS workspace storage: {config.media_gcs_bucket_name}"
-                    )
-                    _workspace_storage = GCSWorkspaceStorage(
-                        config.media_gcs_bucket_name
-                    )
-                else:
-                    storage_dir = (
-                        config.workspace_storage_dir
-                        if config.workspace_storage_dir
-                        else None
-                    )
-                    logger.info(
-                        f"Using local workspace storage: {storage_dir or 'default'}"
-                    )
-                    _workspace_storage = LocalWorkspaceStorage(storage_dir)
-
-    return _workspace_storage
+    # --- GCS storage (per event loop) ---
+    loop_id = id(asyncio.get_running_loop())
+    if loop_id not in _gcs_storages:
+        logger.info(
+            f"Creating GCS workspace storage for loop {loop_id}: "
+            f"{config.media_gcs_bucket_name}"
+        )
+        _gcs_storages[loop_id] = GCSWorkspaceStorage(config.media_gcs_bucket_name)
+    return _gcs_storages[loop_id]
 
 
 async def shutdown_workspace_storage() -> None:
-    """
-    Properly shutdown the global workspace storage backend.
+    """Shut down workspace storage for the **current** event loop.
 
-    Closes aiohttp sessions and other resources for GCS backend.
-    Should be called during application shutdown.
+    Closes the ``aiohttp`` session owned by the current loop's GCS instance.
+    Each worker thread should call this on its own loop before the loop is
+    destroyed.  The REST API lifespan hook calls it for the main server loop.
     """
-    global _workspace_storage
+    global _local_storage
 
-    if _workspace_storage is not None:
-        async with _storage_lock:
-            if _workspace_storage is not None:
-                if isinstance(_workspace_storage, GCSWorkspaceStorage):
-                    await _workspace_storage.close()
-                _workspace_storage = None
+    loop_id = id(asyncio.get_running_loop())
+    storage = _gcs_storages.pop(loop_id, None)
+    if storage is not None:
+        await storage.close()
+
+    # Clear local storage only when the last GCS instance is gone
+    # (i.e. full shutdown, not just a single worker stopping).
+    if not _gcs_storages:
+        _local_storage = None
 
 
 def compute_file_checksum(content: bytes) -> str:

autogpt_platform/frontend/src/app/(platform)/build/components/FlowEditor/nodes/CustomNode/components/NodeAdvancedToggle.tsx

@@ -1,21 +1,39 @@
 import { useNodeStore } from "@/app/(platform)/build/stores/nodeStore";
-import { Switch } from "@/components/atoms/Switch/Switch";
+import { Button } from "@/components/atoms/Button/Button";
 import { Text } from "@/components/atoms/Text/Text";
+import { CaretDownIcon } from "@phosphor-icons/react";
 
-export const NodeAdvancedToggle = ({ nodeId }: { nodeId: string }) => {
+type Props = {
+  nodeId: string;
+};
+
+export function NodeAdvancedToggle({ nodeId }: Props) {
   const showAdvanced = useNodeStore(
     (state) => state.nodeAdvancedStates[nodeId] || false,
   );
   const setShowAdvanced = useNodeStore((state) => state.setShowAdvanced);
   return (
-    <div className="flex items-center justify-between gap-2 rounded-b-xlarge border-t border-zinc-200 bg-white px-5 py-3.5">
-      <Text variant="body" className="font-medium text-slate-700">
-        Advanced
-      </Text>
-      <Switch
-        onCheckedChange={(checked) => setShowAdvanced(nodeId, checked)}
-        checked={showAdvanced}
-      />
+    <div className="flex items-center justify-start gap-2 bg-white px-5 pb-3.5">
+      <Button
+        variant="ghost"
+        className="h-fit min-w-0 p-0 hover:border-transparent hover:bg-transparent"
+        onClick={() => setShowAdvanced(nodeId, !showAdvanced)}
+        aria-expanded={showAdvanced}
+      >
+        <Text
+          variant="body"
+          as="span"
+          className="flex items-center gap-2 !font-semibold text-slate-700"
+        >
+          Advanced{" "}
+          <CaretDownIcon
+            size={16}
+            weight="bold"
+            className={`transition-transform ${showAdvanced ? "rotate-180" : ""}`}
+            aria-hidden
+          />
+        </Text>
+      </Button>
     </div>
   );
-};
+}

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

@@ -23,6 +23,7 @@ export function CopilotPage() {
     status,
     error,
     stop,
+    isReconnecting,
     createSession,
     onSend,
     isLoadingSession,
@@ -71,6 +72,7 @@ export function CopilotPage() {
             sessionId={sessionId}
             isLoadingSession={isLoadingSession}
             isCreatingSession={isCreatingSession}
+            isReconnecting={isReconnecting}
             onCreateSession={createSession}
             onSend={onSend}
             onStop={stop}

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

@@ -14,6 +14,8 @@ export interface ChatContainerProps {
   sessionId: string | null;
   isLoadingSession: boolean;
   isCreatingSession: boolean;
+  /** True when backend has an active stream but we haven't reconnected yet. */
+  isReconnecting?: boolean;
   onCreateSession: () => void | Promise<string>;
   onSend: (message: string) => void | Promise<void>;
   onStop: () => void;
@@ -26,11 +28,13 @@ export const ChatContainer = ({
   sessionId,
   isLoadingSession,
   isCreatingSession,
+  isReconnecting,
   onCreateSession,
   onSend,
   onStop,
   headerSlot,
 }: ChatContainerProps) => {
+  const isBusy = status === "streaming" || !!isReconnecting;
   const inputLayoutId = "copilot-2-chat-input";
 
   return (
@@ -56,8 +60,8 @@ export const ChatContainer = ({
                 <ChatInput
                   inputId="chat-input-session"
                   onSend={onSend}
-                  disabled={status === "streaming"}
-                  isStreaming={status === "streaming"}
+                  disabled={isBusy}
+                  isStreaming={isBusy}
                   onStop={onStop}
                   placeholder="What else can I help with?"
                 />

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

@@ -1,63 +1,713 @@
 "use client";
 
+import React from "react";
 import { ToolUIPart } from "ai";
-import { GearIcon } from "@phosphor-icons/react";
+import {
+  CheckCircleIcon,
+  CircleDashedIcon,
+  CircleIcon,
+  FileIcon,
+  FilesIcon,
+  GearIcon,
+  GlobeIcon,
+  ListChecksIcon,
+  MagnifyingGlassIcon,
+  PencilSimpleIcon,
+  TerminalIcon,
+  TrashIcon,
+  WarningDiamondIcon,
+} from "@phosphor-icons/react";
 import { MorphingTextAnimation } from "../../components/MorphingTextAnimation/MorphingTextAnimation";
+import { ToolAccordion } from "../../components/ToolAccordion/ToolAccordion";
+import {
+  ContentCodeBlock,
+  ContentMessage,
+} from "../../components/ToolAccordion/AccordionContent";
+import { OrbitLoader } from "../../components/OrbitLoader/OrbitLoader";
 
 interface Props {
   part: ToolUIPart;
 }
 
+/* ------------------------------------------------------------------ */
+/*  Tool name helpers                                                  */
+/* ------------------------------------------------------------------ */
+
 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));
+/* ------------------------------------------------------------------ */
+/*  Tool categorization                                                */
+/* ------------------------------------------------------------------ */
 
-  switch (part.state) {
-    case "input-streaming":
-    case "input-available":
-      return `Running ${label}…`;
-    case "output-available":
-      return `${label} completed`;
-    case "output-error":
-      return `${label} failed`;
+type ToolCategory =
+  | "bash"
+  | "web"
+  | "file-read"
+  | "file-write"
+  | "file-delete"
+  | "file-list"
+  | "search"
+  | "edit"
+  | "todo"
+  | "other";
+
+function getToolCategory(toolName: string): ToolCategory {
+  switch (toolName) {
+    case "bash_exec":
+      return "bash";
+    case "web_fetch":
+    case "WebSearch":
+    case "WebFetch":
+      return "web";
+    case "read_workspace_file":
+    case "Read":
+      return "file-read";
+    case "write_workspace_file":
+    case "Write":
+      return "file-write";
+    case "delete_workspace_file":
+      return "file-delete";
+    case "list_workspace_files":
+    case "Glob":
+      return "file-list";
+    case "Grep":
+      return "search";
+    case "Edit":
+      return "edit";
+    case "TodoWrite":
+      return "todo";
     default:
-      return `Running ${label}…`;
+      return "other";
   }
 }
 
+/* ------------------------------------------------------------------ */
+/*  Tool icon                                                          */
+/* ------------------------------------------------------------------ */
+
+function ToolIcon({
+  category,
+  isStreaming,
+  isError,
+}: {
+  category: ToolCategory;
+  isStreaming: boolean;
+  isError: boolean;
+}) {
+  if (isError) {
+    return (
+      <WarningDiamondIcon size={14} weight="regular" className="text-red-500" />
+    );
+  }
+  if (isStreaming) {
+    return <OrbitLoader size={14} />;
+  }
+
+  const iconClass = "text-neutral-400";
+  switch (category) {
+    case "bash":
+      return <TerminalIcon size={14} weight="regular" className={iconClass} />;
+    case "web":
+      return <GlobeIcon size={14} weight="regular" className={iconClass} />;
+    case "file-read":
+      return <FileIcon size={14} weight="regular" className={iconClass} />;
+    case "file-write":
+      return <FileIcon size={14} weight="regular" className={iconClass} />;
+    case "file-delete":
+      return <TrashIcon size={14} weight="regular" className={iconClass} />;
+    case "file-list":
+      return <FilesIcon size={14} weight="regular" className={iconClass} />;
+    case "search":
+      return (
+        <MagnifyingGlassIcon size={14} weight="regular" className={iconClass} />
+      );
+    case "edit":
+      return (
+        <PencilSimpleIcon size={14} weight="regular" className={iconClass} />
+      );
+    case "todo":
+      return (
+        <ListChecksIcon size={14} weight="regular" className={iconClass} />
+      );
+    default:
+      return <GearIcon size={14} weight="regular" className={iconClass} />;
+  }
+}
+
+/* ------------------------------------------------------------------ */
+/*  Accordion icon (larger, for the accordion header)                  */
+/* ------------------------------------------------------------------ */
+
+function AccordionIcon({ category }: { category: ToolCategory }) {
+  switch (category) {
+    case "bash":
+      return <TerminalIcon size={32} weight="light" />;
+    case "web":
+      return <GlobeIcon size={32} weight="light" />;
+    case "file-read":
+    case "file-write":
+      return <FileIcon size={32} weight="light" />;
+    case "file-delete":
+      return <TrashIcon size={32} weight="light" />;
+    case "file-list":
+      return <FilesIcon size={32} weight="light" />;
+    case "search":
+      return <MagnifyingGlassIcon size={32} weight="light" />;
+    case "edit":
+      return <PencilSimpleIcon size={32} weight="light" />;
+    case "todo":
+      return <ListChecksIcon size={32} weight="light" />;
+    default:
+      return <GearIcon size={32} weight="light" />;
+  }
+}
+
+/* ------------------------------------------------------------------ */
+/*  Input extraction                                                   */
+/* ------------------------------------------------------------------ */
+
+function getInputSummary(toolName: string, input: unknown): string | null {
+  if (!input || typeof input !== "object") return null;
+  const inp = input as Record<string, unknown>;
+
+  switch (toolName) {
+    case "bash_exec":
+      return typeof inp.command === "string" ? inp.command : null;
+    case "web_fetch":
+    case "WebFetch":
+      return typeof inp.url === "string" ? inp.url : null;
+    case "WebSearch":
+      return typeof inp.query === "string" ? inp.query : null;
+    case "read_workspace_file":
+    case "Read":
+      return (
+        (typeof inp.file_path === "string" ? inp.file_path : null) ??
+        (typeof inp.path === "string" ? inp.path : null)
+      );
+    case "write_workspace_file":
+    case "Write":
+      return (
+        (typeof inp.file_path === "string" ? inp.file_path : null) ??
+        (typeof inp.path === "string" ? inp.path : null)
+      );
+    case "delete_workspace_file":
+      return typeof inp.file_path === "string" ? inp.file_path : null;
+    case "Glob":
+      return typeof inp.pattern === "string" ? inp.pattern : null;
+    case "Grep":
+      return typeof inp.pattern === "string" ? inp.pattern : null;
+    case "Edit":
+      return typeof inp.file_path === "string" ? inp.file_path : null;
+    case "TodoWrite": {
+      // Extract the in-progress task name for the status line
+      const todos = Array.isArray(inp.todos) ? inp.todos : [];
+      const active = todos.find(
+        (t: Record<string, unknown>) => t.status === "in_progress",
+      );
+      if (active && typeof active.activeForm === "string")
+        return active.activeForm;
+      if (active && typeof active.content === "string") return active.content;
+      return null;
+    }
+    default:
+      return null;
+  }
+}
+
+function truncate(text: string, maxLen: number): string {
+  if (text.length <= maxLen) return text;
+  return text.slice(0, maxLen).trimEnd() + "…";
+}
+
+/* ------------------------------------------------------------------ */
+/*  Animation text                                                     */
+/* ------------------------------------------------------------------ */
+
+function getAnimationText(part: ToolUIPart, category: ToolCategory): string {
+  const toolName = extractToolName(part);
+  const summary = getInputSummary(toolName, part.input);
+  const shortSummary = summary ? truncate(summary, 60) : null;
+
+  switch (part.state) {
+    case "input-streaming":
+    case "input-available": {
+      switch (category) {
+        case "bash":
+          return shortSummary ? `Running: ${shortSummary}` : "Running command…";
+        case "web":
+          if (toolName === "WebSearch") {
+            return shortSummary
+              ? `Searching "${shortSummary}"`
+              : "Searching the web…";
+          }
+          return shortSummary
+            ? `Fetching ${shortSummary}`
+            : "Fetching web content…";
+        case "file-read":
+          return shortSummary ? `Reading ${shortSummary}` : "Reading file…";
+        case "file-write":
+          return shortSummary ? `Writing ${shortSummary}` : "Writing file…";
+        case "file-delete":
+          return shortSummary ? `Deleting ${shortSummary}` : "Deleting file…";
+        case "file-list":
+          return shortSummary ? `Listing ${shortSummary}` : "Listing files…";
+        case "search":
+          return shortSummary
+            ? `Searching for "${shortSummary}"`
+            : "Searching…";
+        case "edit":
+          return shortSummary ? `Editing ${shortSummary}` : "Editing file…";
+        case "todo":
+          return shortSummary ? `${shortSummary}` : "Updating task list…";
+        default:
+          return `Running ${formatToolName(toolName)}…`;
+      }
+    }
+    case "output-available": {
+      switch (category) {
+        case "bash": {
+          const exitCode = getExitCode(part.output);
+          if (exitCode !== null && exitCode !== 0) {
+            return `Command exited with code ${exitCode}`;
+          }
+          return shortSummary ? `Ran: ${shortSummary}` : "Command completed";
+        }
+        case "web":
+          if (toolName === "WebSearch") {
+            return shortSummary
+              ? `Searched "${shortSummary}"`
+              : "Web search completed";
+          }
+          return shortSummary
+            ? `Fetched ${shortSummary}`
+            : "Fetched web content";
+        case "file-read":
+          return shortSummary ? `Read ${shortSummary}` : "File read completed";
+        case "file-write":
+          return shortSummary ? `Wrote ${shortSummary}` : "File written";
+        case "file-delete":
+          return shortSummary ? `Deleted ${shortSummary}` : "File deleted";
+        case "file-list":
+          return "Listed files";
+        case "search":
+          return shortSummary
+            ? `Searched for "${shortSummary}"`
+            : "Search completed";
+        case "edit":
+          return shortSummary ? `Edited ${shortSummary}` : "Edit completed";
+        case "todo":
+          return "Updated task list";
+        default:
+          return `${formatToolName(toolName)} completed`;
+      }
+    }
+    case "output-error": {
+      switch (category) {
+        case "bash":
+          return "Command failed";
+        case "web":
+          return toolName === "WebSearch" ? "Search failed" : "Fetch failed";
+        default:
+          return `${formatToolName(toolName)} failed`;
+      }
+    }
+    default:
+      return `Running ${formatToolName(toolName)}…`;
+  }
+}
+
+/* ------------------------------------------------------------------ */
+/*  Output parsing helpers                                             */
+/* ------------------------------------------------------------------ */
+
+function parseOutput(output: unknown): Record<string, unknown> | null {
+  if (!output) return null;
+  if (typeof output === "object") return output as Record<string, unknown>;
+  if (typeof output === "string") {
+    const trimmed = output.trim();
+    if (!trimmed) return null;
+    try {
+      const parsed = JSON.parse(trimmed);
+      if (
+        typeof parsed === "object" &&
+        parsed !== null &&
+        !Array.isArray(parsed)
+      )
+        return parsed;
+    } catch {
+      // Return as a message wrapper for plain text output
+      return { _raw: trimmed };
+    }
+  }
+  return null;
+}
+
+/**
+ * Extract text from MCP-style content blocks.
+ * SDK built-in tools (WebSearch, etc.) may return `{content: [{type:"text", text:"..."}]}`.
+ */
+function extractMcpText(output: Record<string, unknown>): string | null {
+  if (Array.isArray(output.content)) {
+    const texts = (output.content as Array<Record<string, unknown>>)
+      .filter((b) => b.type === "text" && typeof b.text === "string")
+      .map((b) => b.text as string);
+    if (texts.length > 0) return texts.join("\n");
+  }
+  return null;
+}
+
+function getExitCode(output: unknown): number | null {
+  const parsed = parseOutput(output);
+  if (!parsed) return null;
+  if (typeof parsed.exit_code === "number") return parsed.exit_code;
+  return null;
+}
+
+function getStringField(
+  obj: Record<string, unknown>,
+  ...keys: string[]
+): string | null {
+  for (const key of keys) {
+    if (typeof obj[key] === "string" && obj[key].length > 0)
+      return obj[key] as string;
+  }
+  return null;
+}
+
+/* ------------------------------------------------------------------ */
+/*  Accordion content per tool category                                */
+/* ------------------------------------------------------------------ */
+
+interface AccordionData {
+  title: string;
+  description?: string;
+  content: React.ReactNode;
+}
+
+function getBashAccordionData(
+  input: unknown,
+  output: Record<string, unknown>,
+): AccordionData {
+  const inp = (input && typeof input === "object" ? input : {}) as Record<
+    string,
+    unknown
+  >;
+  const command = typeof inp.command === "string" ? inp.command : "Command";
+
+  const stdout = getStringField(output, "stdout");
+  const stderr = getStringField(output, "stderr");
+  const exitCode =
+    typeof output.exit_code === "number" ? output.exit_code : null;
+  const timedOut = output.timed_out === true;
+  const message = getStringField(output, "message");
+
+  const title = timedOut
+    ? "Command timed out"
+    : exitCode !== null && exitCode !== 0
+      ? `Command failed (exit ${exitCode})`
+      : "Command output";
+
+  return {
+    title,
+    description: truncate(command, 80),
+    content: (
+      <div className="space-y-2">
+        {stdout && (
+          <div>
+            <p className="mb-1 text-xs font-medium text-slate-500">stdout</p>
+            <ContentCodeBlock>{truncate(stdout, 2000)}</ContentCodeBlock>
+          </div>
+        )}
+        {stderr && (
+          <div>
+            <p className="mb-1 text-xs font-medium text-slate-500">stderr</p>
+            <ContentCodeBlock>{truncate(stderr, 1000)}</ContentCodeBlock>
+          </div>
+        )}
+        {!stdout && !stderr && message && (
+          <ContentMessage>{message}</ContentMessage>
+        )}
+      </div>
+    ),
+  };
+}
+
+function getWebAccordionData(
+  input: unknown,
+  output: Record<string, unknown>,
+): AccordionData {
+  const inp = (input && typeof input === "object" ? input : {}) as Record<
+    string,
+    unknown
+  >;
+  const url =
+    getStringField(inp as Record<string, unknown>, "url", "query") ??
+    "Web content";
+
+  // Try direct string fields first, then MCP content blocks, then raw JSON
+  let content = getStringField(output, "content", "text", "_raw");
+  if (!content) content = extractMcpText(output);
+  if (!content) {
+    // Fallback: render the raw JSON so the accordion isn't empty
+    try {
+      const raw = JSON.stringify(output, null, 2);
+      if (raw !== "{}") content = raw;
+    } catch {
+      /* empty */
+    }
+  }
+
+  const statusCode =
+    typeof output.status_code === "number" ? output.status_code : null;
+  const message = getStringField(output, "message");
+
+  return {
+    title: statusCode
+      ? `Response (${statusCode})`
+      : url
+        ? "Web fetch"
+        : "Search results",
+    description: truncate(url, 80),
+    content: content ? (
+      <ContentCodeBlock>{truncate(content, 2000)}</ContentCodeBlock>
+    ) : message ? (
+      <ContentMessage>{message}</ContentMessage>
+    ) : Object.keys(output).length > 0 ? (
+      <ContentCodeBlock>
+        {truncate(JSON.stringify(output, null, 2), 2000)}
+      </ContentCodeBlock>
+    ) : null,
+  };
+}
+
+function getFileAccordionData(
+  input: unknown,
+  output: Record<string, unknown>,
+): AccordionData {
+  const inp = (input && typeof input === "object" ? input : {}) as Record<
+    string,
+    unknown
+  >;
+  const filePath =
+    getStringField(
+      inp as Record<string, unknown>,
+      "file_path",
+      "path",
+      "pattern",
+    ) ?? "File";
+  const content = getStringField(output, "content", "text", "_raw");
+  const message = getStringField(output, "message");
+  // For Glob/list results, try to show file list
+  const files = Array.isArray(output.files)
+    ? output.files.filter((f: unknown): f is string => typeof f === "string")
+    : null;
+
+  return {
+    title: message ?? "File output",
+    description: truncate(filePath, 80),
+    content: (
+      <div className="space-y-2">
+        {content && (
+          <ContentCodeBlock>{truncate(content, 2000)}</ContentCodeBlock>
+        )}
+        {files && files.length > 0 && (
+          <ContentCodeBlock>
+            {truncate(files.join("\n"), 2000)}
+          </ContentCodeBlock>
+        )}
+        {!content && !files && message && (
+          <ContentMessage>{message}</ContentMessage>
+        )}
+      </div>
+    ),
+  };
+}
+
+interface TodoItem {
+  content: string;
+  status: "pending" | "in_progress" | "completed";
+  activeForm?: string;
+}
+
+function getTodoAccordionData(input: unknown): AccordionData {
+  const inp = (input && typeof input === "object" ? input : {}) as Record<
+    string,
+    unknown
+  >;
+  const todos: TodoItem[] = Array.isArray(inp.todos)
+    ? inp.todos.filter(
+        (t: unknown): t is TodoItem =>
+          typeof t === "object" &&
+          t !== null &&
+          typeof (t as TodoItem).content === "string",
+      )
+    : [];
+
+  const completed = todos.filter((t) => t.status === "completed").length;
+  const total = todos.length;
+
+  return {
+    title: "Task list",
+    description: `${completed}/${total} completed`,
+    content: (
+      <div className="space-y-1 py-1">
+        {todos.map((todo, i) => (
+          <div key={i} className="flex items-start gap-2 text-xs">
+            <span className="mt-0.5 flex-shrink-0">
+              {todo.status === "completed" ? (
+                <CheckCircleIcon
+                  size={14}
+                  weight="fill"
+                  className="text-green-500"
+                />
+              ) : todo.status === "in_progress" ? (
+                <CircleDashedIcon
+                  size={14}
+                  weight="bold"
+                  className="text-blue-500"
+                />
+              ) : (
+                <CircleIcon
+                  size={14}
+                  weight="regular"
+                  className="text-neutral-400"
+                />
+              )}
+            </span>
+            <span
+              className={
+                todo.status === "completed"
+                  ? "text-muted-foreground line-through"
+                  : todo.status === "in_progress"
+                    ? "font-medium text-foreground"
+                    : "text-muted-foreground"
+              }
+            >
+              {todo.content}
+            </span>
+          </div>
+        ))}
+      </div>
+    ),
+  };
+}
+
+function getDefaultAccordionData(
+  output: Record<string, unknown>,
+): AccordionData {
+  const message = getStringField(output, "message");
+  const raw = output._raw;
+  const mcpText = extractMcpText(output);
+
+  let displayContent: string;
+  if (typeof raw === "string") {
+    displayContent = raw;
+  } else if (mcpText) {
+    displayContent = mcpText;
+  } else if (message) {
+    displayContent = message;
+  } else {
+    try {
+      displayContent = JSON.stringify(output, null, 2);
+    } catch {
+      displayContent = String(output);
+    }
+  }
+
+  return {
+    title: "Output",
+    description: message ?? undefined,
+    content: (
+      <ContentCodeBlock>{truncate(displayContent, 2000)}</ContentCodeBlock>
+    ),
+  };
+}
+
+function getAccordionData(
+  category: ToolCategory,
+  input: unknown,
+  output: Record<string, unknown>,
+): AccordionData {
+  switch (category) {
+    case "bash":
+      return getBashAccordionData(input, output);
+    case "web":
+      return getWebAccordionData(input, output);
+    case "file-read":
+    case "file-write":
+    case "file-delete":
+    case "file-list":
+    case "search":
+    case "edit":
+      return getFileAccordionData(input, output);
+    case "todo":
+      return getTodoAccordionData(input);
+    default:
+      return getDefaultAccordionData(output);
+  }
+}
+
+/* ------------------------------------------------------------------ */
+/*  Component                                                          */
+/* ------------------------------------------------------------------ */
+
 export function GenericTool({ part }: Props) {
+  const toolName = extractToolName(part);
+  const category = getToolCategory(toolName);
   const isStreaming =
     part.state === "input-streaming" || part.state === "input-available";
   const isError = part.state === "output-error";
+  const text = getAnimationText(part, category);
+
+  const output = parseOutput(part.output);
+  const hasOutput =
+    part.state === "output-available" &&
+    !!output &&
+    Object.keys(output).length > 0;
+  const hasError = isError && !!output;
+
+  // TodoWrite: always show accordion from input (the todo list lives in input)
+  const hasTodoInput =
+    category === "todo" &&
+    part.input &&
+    typeof part.input === "object" &&
+    Array.isArray((part.input as Record<string, unknown>).todos);
+
+  const showAccordion = hasOutput || hasError || hasTodoInput;
+  const accordionData = showAccordion
+    ? getAccordionData(category, part.input, output ?? {})
+    : null;
 
   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"
-          }
+        <ToolIcon
+          category={category}
+          isStreaming={isStreaming}
+          isError={isError}
         />
         <MorphingTextAnimation
-          text={getAnimationText(part)}
+          text={text}
           className={isError ? "text-red-500" : undefined}
         />
       </div>
+
+      {showAccordion && accordionData ? (
+        <ToolAccordion
+          icon={<AccordionIcon category={category} />}
+          title={accordionData.title}
+          description={accordionData.description}
+          titleClassName={isError ? "text-red-500" : undefined}
+        >
+          {accordionData.content}
+        </ToolAccordion>
+      ) : null}
     </div>
   );
 }

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

@@ -50,6 +50,14 @@ export function useChatSession() {
     );
   }, [sessionQuery.data, sessionId]);
 
+  // Expose active_stream info so the caller can trigger manual resume
+  // after hydration completes (rather than relying on AI SDK's built-in
+  // resume which fires before hydration).
+  const hasActiveStream = useMemo(() => {
+    if (sessionQuery.data?.status !== 200) return false;
+    return !!sessionQuery.data.data.active_stream;
+  }, [sessionQuery.data]);
+
   const { mutateAsync: createSessionMutation, isPending: isCreatingSession } =
     usePostV2CreateSession({
       mutation: {
@@ -102,6 +110,7 @@ export function useChatSession() {
     sessionId,
     setSessionId,
     hydratedMessages,
+    hasActiveStream,
     isLoadingSession: sessionQuery.isLoading,
     createSession,
     isCreatingSession,

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

@@ -29,6 +29,7 @@ export function useCopilotPage() {
     sessionId,
     setSessionId,
     hydratedMessages,
+    hasActiveStream,
     isLoadingSession,
     createSession,
     isCreatingSession,
@@ -80,14 +81,31 @@ export function useCopilotPage() {
                 },
               };
             },
+            // Resume: GET goes to the same URL as POST (backend uses
+            // method to distinguish).  Override the default formula which
+            // would append /{chatId}/stream to the existing path.
+            prepareReconnectToStreamRequest: () => ({
+              api: `/api/chat/sessions/${sessionId}/stream`,
+            }),
           })
         : null,
     [sessionId],
   );
 
-  const { messages, sendMessage, stop, status, error, setMessages } = useChat({
+  const {
+    messages,
+    sendMessage,
+    stop,
+    status,
+    error,
+    setMessages,
+    resumeStream,
+  } = useChat({
     id: sessionId ?? undefined,
     transport: transport ?? undefined,
+    // Don't use resume: true — it fires before hydration completes, causing
+    // the hydrated messages to overwrite the resumed stream.  Instead we
+    // call resumeStream() manually after hydration + active_stream detection.
   });
 
   // Abort the stream if the backend doesn't start sending data within 12s.
@@ -108,13 +126,31 @@ export function useCopilotPage() {
     return () => clearTimeout(timer);
   }, [status]);
 
+  // Hydrate messages from the REST session endpoint.
+  // Skip hydration while streaming to avoid overwriting the live stream.
   useEffect(() => {
     if (!hydratedMessages || hydratedMessages.length === 0) return;
+    if (status === "streaming" || status === "submitted") return;
     setMessages((prev) => {
       if (prev.length >= hydratedMessages.length) return prev;
       return hydratedMessages;
     });
-  }, [hydratedMessages, setMessages]);
+  }, [hydratedMessages, setMessages, status]);
+
+  // Resume an active stream AFTER hydration completes.
+  // The backend returns active_stream info when a task is still running.
+  // We wait for hydration so the AI SDK has the conversation history
+  // before the resumed stream appends the in-progress assistant message.
+  const hasResumedRef = useRef<string | null>(null);
+  useEffect(() => {
+    if (!hasActiveStream || !sessionId) return;
+    if (!hydratedMessages || hydratedMessages.length === 0) return;
+    if (status === "streaming" || status === "submitted") return;
+    // Only resume once per session to avoid re-triggering after stream ends
+    if (hasResumedRef.current === sessionId) return;
+    hasResumedRef.current = sessionId;
+    resumeStream();
+  }, [hasActiveStream, sessionId, hydratedMessages, status, resumeStream]);
 
   // Poll session endpoint when a long-running tool (create_agent, edit_agent)
   // is in progress. When the backend completes, the session data will contain
@@ -197,12 +233,18 @@ export function useCopilotPage() {
     }
   }, [isDeleting]);
 
+  // True while we know the backend has an active stream but haven't
+  // reconnected yet.  Used to disable the send button and show stop UI.
+  const isReconnecting =
+    hasActiveStream && status !== "streaming" && status !== "submitted";
+
   return {
     sessionId,
     messages,
     status,
     error,
     stop,
+    isReconnecting,
     isLoadingSession,
     isCreatingSession,
     isUserLoading,