[BUG] Cowork Windows 10 22H2 x64: `sandbox-helper: no Plan9 drive shares mounted` — device_bash dead since 2026-09-08, survives app restart and app update

### Preflight Checklist

- [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Abug) and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code

### What's Wrong?

Every device_bash call from a Cowork session on this machine fails immediately,
including `echo ok` - a command that needs no mounted folder at all:

sandbox-helper: no Plan9 drive shares mounted under /mnt/.virtiofs-root/shared

The sandbox VM's shell never starts, so nothing can be run on the computer any
more: no scripts, no in-place file edits, no renames or moves.

The rest of the device bridge is healthy - get_device_info, device_list_dir,
device_stage_files and device_commit_files all still work normally. The failure
is isolated to attaching the Plan9 shares to the guest VM.

This worked fine on 2026-09-04 and broke on 2026-09-08, the day Windows 10
cumulative update KB5122878 (OS build 19045.7725) installed itself. It survives
a full app restart and an app version upgrade (1.46388.2 -> 1.49585.0), so the
change came from the host OS, not from the app.

Same error string as #92958, but that report is Windows 11 ARM64 with a
different KB (KB5124012). This is Windows 10 22H2 x64.

### What Should Happen?

device_bash should start and run commands as it did until 2026-09-04: the
connected folders mounted under $HOME/mnt/, and plain commands like `echo ok`
succeeding regardless of mount state.

At minimum, if the host cannot attach the Plan9 shares, the failure should be
reported where it actually happens - host-side share attachment returned zero
shares - instead of only as a guest-side mount error that gives no hint that a
Windows update is involved.

### Error Messages/Logs

```shell
sandbox-helper: no Plan9 drive shares mounted under /mnt/.virtiofs-root/shared

[Note: mnt/Email processing, mnt/_Inbox, mnt/HRC, mnt/RenDit, mnt/Claude skills
failed to mount and cannot be reached from this shell. Those connected folders are
still reachable via device_list_dir / device_stage_files / device_commit_files,
by their paths on this device.]
```

### Steps to Reproduce

1. On Windows 10 22H2 x64 with KB5122878 installed (OS build 19045.7725),
open the Claude desktop app and start a Cowork task.

2. Connect one or more local folders to that task. In my case five folders
under C:\Users\\Dropbox\ — the specific folders do not matter.

3. Ask Claude to run a shell command on the computer, e.g.
"run `echo ok` on my computer". This invokes the device_bash tool.

4. Every device_bash call fails immediately with the error below — including
`echo ok`, which needs no mounted folder at all. The shell never starts,
so this is not a per-command or per-path failure.

5. Confirm the rest of the device bridge is unaffected: ask Claude to list one
of the connected folders (device_list_dir). That succeeds and returns the
folder contents normally. device_stage_files and device_commit_files also
still work.

6. Fully quit the desktop app and relaunch it, then repeat step 3.
Identical failure.

Regression window: device_bash ran dozens of commands successfully against
all five folders on 2026-09-04 (ls, cp, mv, md5sum, python3 scripts, no
errors). It was broken on 2026-09-08 and has stayed broken since. KB5122878
installed itself on 2026-09-08. Nothing else changed on the machine.

### Claude Model

Opus

### Is this a regression?

Yes, this worked in a previous version

### Last Working Version

_No response_

### Claude Code Version

1.49585.0 (Claude desktop app / Cowork — Claude Code CLI not installed; also failed on 1.46388.2)

### Platform

Anthropic API

### Operating System

Windows

### Terminal/Shell

Other

### Additional Information

## Why "Last Working Version" is misleading here

This is a regression, but not an app-version regression. The same app build
that worked on 2026-09-04 (1.46388.2) was failing on 2026-09-08. The app has
since auto-updated to 1.49585.0 and the failure is byte-identical. So please
do not bisect on app versions — the change came from the host OS.

## Environment

| | |
|---|---|
| Platform | win32 |
| Arch | x64 |
| OS | Windows 10 22H2, build 19045.7725 |
| KB5122878 | installed 2026-09-08 by NT AUTHORITY\SYSTEM (automatic Windows Update) |
| Claude desktop app | 1.49585.0 (also failed on 1.46388.2) |
| Electron | 44.2.0 |
| Node | 24.20.0 |
| Connected folders | 5, all under C:\Users\\Dropbox\ (Dropbox-synced, and working fine until 2026-09-08) |

Verified on the machine (PowerShell):

OS: 10.0.19045.7725

HotFixID Description InstalledBy InstalledOn
-------- ----------- ----------- -----------
KB5122878 Security Update NT AUTHORITY\SYSTEM 8-9-2026 00:00:00

The same query also asked for KB5124012 and returned nothing: that KB is NOT
installed here — it targets Windows 11 (build 28000.x) and does not apply to
this machine.

## Not a terminal issue

The "Terminal/Shell" field does not really apply: there is no terminal
involved. The failing shell is Cowork's own sandboxed bash running on the
device (the device_bash tool). PowerShell was only used to collect the OS
build and hotfix evidence above.

## Scope: host-side bridge is healthy, guest VM is not

Everything that does not go through the sandbox VM keeps working:

- get_device_info returns all five connected folders correctly
- device_list_dir returns folder contents normally
- device_stage_files / device_commit_files both work

The failure is isolated to attaching the Plan9 shares to the guest VM.

## Relation to #92958

#92958 reports the same error string —
"sandbox-helper: no Plan9 drive shares mounted under /mnt/.virtiofs-root/shared"
— on Windows 11 ARM64, and traces it by A/B rollback to KB5124012
(build 28000.2954), where add_plan9_shares logs "completed" while attaching
zero shares.

This report is Windows 10 22H2 x64: a different OS, a different architecture
and a different KB, with the same guest-side symptom and the same break date
(September 2026 Patch Tuesday). That suggests the September 2026 cumulative
updates broke Plan9 share attachment more broadly, rather than #92958 being
an ARM64-specific defect.

To be explicit about evidence quality: I have not done an A/B rollback, so on
this machine the KB link is correlation. The window is one day wide and
nothing else changed.

## Impact

Cowork sessions can still read and write files through the staging tools, but
cannot run anything on the machine — no scripts, no in-place edits, no renames
or moves. For file-organisation workflows that removes most of the value:
every operation becomes stage-to-cloud, process, commit-back, with renames and
moves left to the user by hand.

## Asks

1. Is the add_plan9_shares silent-failure path in #92958 the same root cause
here, or a separate Windows 10 issue?
2. Is there a workaround short of uninstalling a security update? Windows 10
is on ESU now, so rolling back a cumulative update is a poor trade.
3. The error message gives no hint that a Windows update is involved.
Surfacing "host share attachment returned zero shares" at the point of
failure, instead of only the guest-side mount error, would make this
diagnosable.

Start with the device_bash failure and the add_plan9_shares path discussed in #92958, then reproduce on Windows 10 22H2 x64 with KB5122878 while checking whether host-side attachment returns zero shares. Compare the host result with the guest error; done means identifying this Windows 10 path and reporting the failure at the host-side attachment point rather than only the guest mount message.

Do mô hình lập chỉ mục viết ra từ nội dung của issue.