Where does the bug appear (feature/product)?
Cursor IDE
Describe the Bug
Hi,
I’m experiencing a reproducible Remote SSH Extension Host crash loop with Cursor 3.21.16 on a managed Debian 12 server.
The SSH connection itself works and the Cursor Server starts successfully. However, the Remote Extension Host is killed approximately 30 seconds after startup. Cursor automatically restarts it, and after three crashes Cursor shows:
The remote extension host terminated unexpectedly 3 times within the last 5 minutes.
Steps to Reproduce
Environment
Local
- macOS
- Cursor 3.21.16
- Connection via Cursor Remote SSH
Remote
- Managed hosting environment
- No root access
- Debian GNU/Linux 12 (bookworm)
- x86_64
- Kernel:
Linux XXX.your-server.de 6.1.0-53-amd64 #1 SMP PREEMPT_DYNAMIC Debian 6.1.187-1 (2026-09-07) x86_64 GNU/Linux
User namespaces are disabled by the hosting provider:
$ cat /proc/sys/user/max_user_namespaces
0
The hosting provider has confirmed that user namespaces cannot be enabled in this environment.
Trying to create a user namespace manually:
$ unshare --user --map-root-user true
unshare: unshare failed: No space left on device
$ echo $?
1
The “No space left on device” message is not caused by actual disk or memory exhaustion. User namespaces are explicitly disabled (max_user_namespaces=0).
Memory / OOM checks
The server has plenty of available memory:
total used free shared buff/cache available
Mem: 251Gi 14Gi 6.4Gi 182Mi 232Gi 237Gi
Swap: 4.0Gi 731Mi 3.3Gi
The process is not being killed by the cgroup OOM killer:
memory.max:
max
memory.events:
low 0
high 0
max 0
oom 0
oom_kill 0
oom_group_kill 0
The process belongs to:
0::/system.slice/ssh.service
Relevant process limits:
Max processes 2048 2048
Max open files 1024 1048576
Max resident set unlimited unlimited
Max address space unlimited unlimited
So I don’t see evidence of a memory/OOM-related SIGKILL.
Cursor Remote logs
The Extension Host initially starts normally:
2026-09-21 10:27:08.571 [info] [127.0.0.1][cbf22b78][ExtensionHostConnection] <58667> Launched Extension Host Process.
Shortly afterwards, the Remote Extension Host starts detecting sandbox support:
[ExtHostShellExec] Detecting sandbox support...
[ExtHostShellExec] Platform: linux, Shell type: unknown
[ExtHostShellExec] Sandbox helper path:
/usr/home/planrt/.cursor-server/bin/linux-x64/8ae78e8eee1e63479c7e0504b664bc0a80c68000/resources/helpers/cursorsandbox
[ExtHostShellExec] Probing sandbox support...
[shell-exec:sandbox] [sandboxPreflight] Running preflight with binary:
.../resources/helpers/cursorsandbox
The sandbox preflight then times out after exactly 15 seconds:
2026-09-21 10:27:25.038 [error] [shell-exec:sandbox] [primeSandboxHelperSupport] Preflight timed out after 15000ms, treating sandbox as unsupported
2026-09-21 10:27:25.038 [info] [ExtHostShellExec] Sandbox support detected: false, platform: linux, kernelVersion: 6.1.0-53-amd64
About 12 seconds later, the Extension Host receives SIGKILL:
2026-09-21 10:27:37.182 [info] [127.0.0.1][cbf22b78][ExtensionHostConnection] <58667> Extension Host Process exited with code: null, signal: SIGKILL.
The File Watcher is also killed:
2026-09-21 10:27:37.240 [error] [File Watcher (universal)] restarting watcher after unexpected error: terminated by itself with code null, signal: SIGKILL (ETERM)
Cursor immediately launches another Extension Host:
2026-09-21 10:27:37.462 [info] [127.0.0.1][d5804022][ExtensionHostConnection] <59140> Launched Extension Host Process.
The exact same sequence happens again:
2026-09-21 10:27:53.899 [error] [shell-exec:sandbox] [primeSandboxHelperSupport] Preflight timed out after 15000ms, treating sandbox as unsupported
2026-09-21 10:27:53.899 [info] [ExtHostShellExec] Sandbox support detected: false, platform: linux, kernelVersion: 6.1.0-53-amd64
followed by:
2026-09-21 10:28:07.222 [info] [127.0.0.1][d5804022][ExtensionHostConnection] <59140> Extension Host Process exited with code: null, signal: SIGKILL.
Cursor launches the Extension Host again as PID 59538 and the same sequence repeats for a third time.
So the behavior is consistently:
Remote Extension Host starts
|
v
Cursor detects sandbox support
|
v
cursorsandbox preflight starts
|
v
preflight hangs for 15 seconds
|
v
"Sandbox support detected: false"
|
v
~12-14 seconds later
|
v
Extension Host receives SIGKILL
|
v
File Watcher also receives SIGKILL
|
v
Cursor restarts Extension Host
|
v
same sequence repeats
Extensions appear to initialize successfully
The Cursor extensions themselves appear to activate successfully before the Extension Host is killed:
Extension activated success: anysphere.cursor-agent-exec
Extension activated success: anysphere.cursor-explorer
Extension activated success: anysphere.cursor-shadow-workspace
Extension activated success: anysphere.cursor-retrieval
There are also messages such as:
No bundle location found for extension anysphere.cursor-agent-exec
No bundle location found for extension anysphere.cursor-checkout
No bundle location found for extension anysphere.cursor-polyfills-remote
However, those extensions subsequently report successful activation, so I’m not sure whether these messages are related.
Attempted workaround: disable sandbox
Since user namespaces cannot be enabled on this server, I tried explicitly disabling Cursor’s sandbox.
I created:
~/.cursor/sandbox.json
with:
{
"type": "insecure_none"
}
I then stopped the existing Cursor Server:
pkill -u "$USER" -f cursor-server
and completely restarted Cursor on the Mac before reconnecting via Remote SSH.
Unfortunately, this does not change the behavior.
Even with:
{
"type": "insecure_none"
}
the Remote Extension Host still executes the cursorsandbox preflight:
Detecting sandbox support...
Probing sandbox support...
Running preflight with binary: .../cursorsandbox
Preflight timed out after 15000ms
Sandbox support detected: false
and the Extension Host is subsequently SIGKILLed again.
Important observation
The server explicitly has:
/proc/sys/user/max_user_namespaces = 0
and therefore:
unshare --user --map-root-user true
fails.
At the same time, Cursor’s Remote Extension Host repeatedly tries to probe cursorsandbox, waits 15 seconds for the probe to time out, correctly determines:
Sandbox support detected: false
but the Extension Host is then SIGKILLed shortly afterwards.
I don’t want to claim that the sandbox preflight itself is definitely causing the SIGKILL, but the sequence is highly reproducible and happens on every Extension Host restart.
There is also no evidence of an OOM kill (oom_kill=0) and memory availability is >200 GiB.
Questions
-
Should Cursor Remote SSH support Linux servers where
max_user_namespaces=0? -
Once Cursor detects:
Sandbox support detected: false
should the Remote Extension Host simply continue without sandboxing?
-
Is there a supported configuration or environment variable to completely disable the
cursorsandboxcapability/preflight check for Remote SSH hosts? -
Should:
{
"type": "insecure_none"
}
prevent this preflight, or does this setting only affect sandboxing of Agent command execution?
-
Could Cursor itself be sending the SIGKILL after the sandbox detection/preflight fails, or is there another Cursor subprocess/watchdog that could explain both the Extension Host and File Watcher receiving SIGKILL?
-
Is this a known regression in Cursor 3.21.x?
Unfortunately, enabling user namespaces is not an option because this is a managed hosting environment and the hosting provider explicitly disables them.
If useful, I can provide the complete:
remoteagent.logremoteexthost.logCursor Agent Exec.log
or run additional diagnostics.
Thanks!
Operating System
Linux
Version Information
3.21.16 on MacOs
Does this stop you from using Cursor
Yes - Cursor is unusable