fix(copilot): sandbox kill, tool event logging, and background task UX

- Fix sandbox process kill: use start_new_session + os.killpg to kill
  the entire bwrap process group on timeout (proc.kill alone only kills
  the parent, leaving children running until natural completion)
- Add StreamToolInputAvailable/StreamToolOutputAvailable to publish_chunk
  logging filter so tool events are visible in Docker logs
- Add system prompt instruction telling Claude not to use
  run_in_background on Task tool (gets denied by security hooks)
- Add tool event debug logging in SDK streaming loop for tracing
  tool execution visibility issues

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

@@ -100,6 +100,8 @@ _SDK_TOOL_SUPPLEMENT = """
 - 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.
+- When using the Task tool, NEVER set `run_in_background` to true.
+  All tasks must run in the foreground.
 """
 
 
@@ -680,6 +682,17 @@ async def stream_chat_completion_sdk(
                         if isinstance(response, StreamStart):
                             continue
 
+                        # Log tool events for debugging visibility issues
+                        if isinstance(
+                            response,
+                            (StreamToolInputAvailable, StreamToolOutputAvailable),
+                        ):
+                            logger.info(
+                                "[SDK] Tool event: %s, tool=%s",
+                                type(response).__name__,
+                                getattr(response, "toolName", "N/A"),
+                            )
+
                         yield response
 
                         if isinstance(response, StreamTextDelta):

autogpt_platform/backend/backend/copilot/stream_registry.py

@@ -227,7 +227,14 @@ async def publish_chunk(
         # Only log timing for significant chunks or slow operations
         if (
             chunk_type
-            in ("StreamStart", "StreamFinish", "StreamTextStart", "StreamTextEnd")
+            in (
+                "StreamStart",
+                "StreamFinish",
+                "StreamTextStart",
+                "StreamTextEnd",
+                "StreamToolInputAvailable",
+                "StreamToolOutputAvailable",
+            )
             or total_time > 50
         ):
             logger.info(

autogpt_platform/backend/backend/copilot/tools/sandbox.py

@@ -13,6 +13,7 @@ import logging
 import os
 import platform
 import shutil
+import signal
 
 logger = logging.getLogger(__name__)
 
@@ -245,6 +246,7 @@ async def run_sandboxed(
             stderr=asyncio.subprocess.PIPE,
             cwd=cwd,
             env=safe_env,
+            start_new_session=True,  # Own process group for clean kill
         )
 
         try:
@@ -255,7 +257,13 @@ async def run_sandboxed(
             stderr = stderr_bytes.decode("utf-8", errors="replace")
             return stdout, stderr, proc.returncode or 0, False
         except asyncio.TimeoutError:
-            proc.kill()
+            # Kill entire process group (bwrap + all children).
+            # proc.kill() alone only kills the bwrap parent, leaving
+            # children running until they finish naturally.
+            try:
+                os.killpg(proc.pid, signal.SIGKILL)
+            except ProcessLookupError:
+                pass  # Already exited
             await proc.communicate()
             return "", f"Execution timed out after {timeout}s", -1, True