Windows: orphaned git.exe is hung pre-main in detect_msys_tty/NtQueryObject; the wedged process is a grandchild of the cmd\git.exe shim and escapes direct-child cleanup #41783

What version of Codex CLI is running?

codex-cli 0.151.0-alpha.7.2, run via codex.exe app-server launched by the Codex Windows desktop app (package OpenAI.Codex_26.825.6671.0_x64).

What subscription do you have?

Omitting; not relevant to this report.

Which model were you using?

Not model-dependent. The leak is in the git probe path.

What platform is your computer?

Windows 11 Pro 10.0.22631 x64. Git for Windows 2.39.2.windows.1 at C:\Program Files\Git.

What issue are you seeing?

Codex leaves git.exe processes alive forever on Windows. Same symptom as #17229 and #29110, but I attached a debugger to the live orphans and got the actual root cause, which as far as I can tell has not been posted before. It is not a polling-frequency problem and it is not a child blocked writing to a stdout pipe.

The hung git is frozen before main() runs.

A non-invasive cdb -p <pid> -pv -c "~*k" attach to six orphans that had been alive 6.8 hours gave the identical stack on every one:

ntdll!NtQueryObject+0x14
git+0x285702
git+0x2864d7
git+0x284647
git+0x12ef      <- wmain
git+0x1406      <- CRT entry
KERNEL32!BaseThreadInitThunk+0x1d
ntdll!RtlUserThreadStart+0x28

Git for Windows ships no PDBs, so I identified the frame from the call site itself:

mov     r9d, 3FEh        ; ResultLength arg == sizeof(buffer) - 2, buffer is BYTE[1024]
lea     r8,  [rsp+40h]   ; ObjectInformation
mov     rcx, rbx         ; Handle
mov     edx, 1           ; ObjectInformationClass = ObjectNameInformation
call    git+0x2b3f80     ; NtQueryObject
test    eax, eax
js      <bail>
cmp     dword ptr [rsp+3Ch], 0Fh

That is verbatim detect_msys_tty() in git-for-windows compat/winansi.c, called from winansi_init() at startup. It runs against fd 0, 1 and 2 whenever the handle is a pipe, to sniff for an MSYS2 pty name. NtQueryObject(ObjectNameInformation) on a wedged pipe handle never returns and the wait is not alertable. So the process never reaches main(), no git-level timeout can apply, no parent-side timeout can reach it once the parent is gone, and it survives indefinitely. Mine were killed manually at 6.8 hours.

The measurable signature, which I think is the useful part. Blocked before its first file read, such a process has done literally nothing, and the counters never move:

ProcessId KernelModeTime UserModeTime ReadOperationCount WriteOperationCount ageMin
    40788              0            0                  0                   0  425.3
    54996              0            0                  0                   0  425.3
    35868              0            0                  0                   0  425.3
     7180              0            0                  0                   0  425.2
    55596              0            0                  0                   0  425.2
    53560              0            0                  0                   0  425.2

24 DLLs were mapped and Ldr.Initialized: Yes, so the loader completed and it is genuinely parked in the kernel inside winansi_init(). For contrast, on the same host at the same moment, a real git reset --hard showed 209 reads / 8538 writes and another concurrent git showed 644 reads, within milliseconds of starting. Zero on all four is therefore a clean discriminator between "hung like this" and "slow but working", and it explains the "0 CPU-time delta" observations several people already reported in #17229.

Attribution. I ran a WMI __InstanceCreationEvent watcher over git.exe for 15 minutes, walking the parent chain of every spawn. 1230 git spawns were observed. Every process carrying this argv shape traced to codex.exe app-server (pid 35984) under ChatGPT.exe under explorer.exe. Zero came from any other agent on the box, and I had several running concurrently.

The six orphans' argv and cwd:

git.exe -c core.hooksPath=NUL -c core.fsmonitor=false status --porcelain     (x3)
git.exe -c core.hooksPath=NUL -c core.fsmonitor=false remote -v              (x1)
git.exe config --null --get core.fsmonitor                                   (x2)
CurrentDirectory: C:\tmp\cwt-chip-<redacted>-20260830\   (all six, a Codex worktree)

core.hooksPath=NUL matches DISABLED_HOOKS_PATH in codex-rs/git-utils/src/info.rs, and config --null --get core.fsmonitor matches codex-rs/git-utils/src/fsmonitor.rs and codex-rs/tui/src/get_git_diff.rs.

Why they end up orphaned, and the part I would like you to check.

On this host git resolves to C:\Program Files\Git\cmd\git.exe, a shim that spawns C:\Program Files\Git\mingw64\bin\git.exe as a child. So Codex's direct child is the shim and the process that hangs is a grandchild. I captured that two-level chain live:

pid=28584  "git" -c safe.bareRepository=explicit -c core.hooksPath=NUL -c core.fsmonitor=false remote get-url origin
    ^ pid=35984 codex.exe ... app-server
pid=37004  git.exe -c safe.bareRepository=explicit -c core.hooksPath=NUL -c core.fsmonitor=false remote get-url origin
    ^ pid=28584 (the shim)

Anything that reaps only the direct child leaves the hung grandchild behind. Being careful about what I have and have not verified:

• codex-rs/git-utils/src/git_process.rs does already do the right thing: JobObject::create() plus spawn_contained, with KillGitProcessTreeOnDrop and a preserve_descendants() call on the success path. I am not claiming that path is broken.
• But spawn_git_command's Err(_) fallback sets command.creation_flags(0) and spawns with job: None, and there is no #[cfg(windows)] impl Drop for KillGitProcessTreeOnDrop. In that state the only reaper left is kill_on_drop(true), which terminates the direct child only, which is exactly the shape I observed. A failed contained spawn silently degrades to leaking the tree.
• Separately, the probes in codex-rs/tui/src/get_git_diff.rs do not go through git_process.rs at all. They build a WorkspaceCommand and run it through AppServerWorkspaceCommandRunner, i.e. the app-server exec path. My orphans' parent was the app-server, and two of the six are exactly that probe (config --null --get core.fsmonitor). I could not verify from source whether the app-server exec path job-contains on Windows, and that is the specific thing I am asking you to check.

I am deliberately not asserting which of those two is the cause. I first assumed "Codex kills the wrapper and not the tree", then read git_process.rs and found that framing wrong for that path, so I am reporting the evidence and the two candidate gaps rather than a conclusion.

What steps can reproduce the bug?

I do not have a deterministic trigger for the NtQueryObject wedge itself; it is a race on the pipe handle, and here it surfaced in bursts (six orphans inside one 5.1 second window, in two ticks about 5 seconds apart). The detection is reliable, so to confirm it on your own machines:

1. Run Codex on Windows against a git repo for a while, ideally with parallel tasks. Other reporters in Codex Windows App keeps spawning git.exe status --porcelain=v1 -z and leaves orphan git.exe / conhost.exe processes #17229 see it worst with subagents.
2. Enumerate hung git with the zero-work predicate:

$now = Get-Date
$all = Get-CimInstance Win32_Process
foreach ($p in ($all | Where-Object { $_.Name -eq 'git.exe' })) {
  $ageMin = [math]::Round(($now - $p.CreationDate).TotalMinutes, 1)
  $zero = ($p.KernelModeTime -eq 0) -and ($p.UserModeTime -eq 0) -and
          ($p.ReadOperationCount -eq 0) -and ($p.WriteOperationCount -eq 0)
  if ($zero -and $ageMin -gt 2) {
    '{0} ppid={1} {2}min {3}' -f $p.ProcessId, $p.ParentProcessId, $ageMin,
      ($p.CommandLine -replace '\s+', ' ')
  }
}

3. Confirm the stack on any hit. This attaches non-invasively and leaves the process running, so it does not destroy the evidence:

"C:\Program Files (x86)\Windows Kits\10\Debuggers\x64\cdb.exe" -p <pid> -pv -c "~*k; q"

ntdll!NtQueryObject should appear four frames above BaseThreadInitThunk.

What is the expected behavior?

A git probe Codex has already given up on should not outlive the session, and certainly not outlive the app. Concretely:

1. Make Windows job containment cover every git spawn, including the app-server / WorkspaceCommand probe path, not just git-utils/src/git_process.rs. Because the real git is a grandchild of the cmd\git.exe shim on a stock Git for Windows install, a job object or an explicit tree kill is the only thing that actually reaps it.
2. Make spawn_git_command's fallback fail closed. Today, JobObject::create() or spawn_contained failing downgrades to an uncontained spawn whose only cleanup is kill_on_drop, which cannot reach the grandchild. Failing the probe outright beats leaking a process that can never exit.
3. Optional, and it would sidestep the git bug entirely: detect_msys_tty only runs for handles whose GetFileType is FILE_TYPE_PIPE. spawn_git_command already passes Stdio::null() for stdin, which is why fd 0 is safe; fd 1 and 2 are piped() and are what hang. For the small existence probes, capturing output via a temp file instead of a pipe would mean git never enters the hanging path. This is a workaround for a Git for Windows bug rather than a Codex bug, so treat it as a suggestion, not the ask.

Additional information

Related: #17229 (same symptom, open since April, 33 comments, no root cause posted; several comments there independently report "all recorded parent PIDs no longer existed" and "0 CPU-time delta", which this stack explains) and #29110.

Reproduced on this host twice, seven days apart: 2026-08-23 (4 orphans, 2.5 to 4 minutes old when found) and 2026-08-30/31 (6 orphans, 6.8 hours old). Both times on trivial commands that finish in well under a second when they run at all.

I mitigated locally with a reaper that kills git.exe only when all four counters are exactly zero, the process is past an age threshold, its parent is gone or its parent PID was recycled, and it has no live child. That last clause matters: a cmd\git.exe shim parked in WaitForSingleObject over a real working git also reads as zero-work, and killing it would orphan live work. I verified the gate by orphaning a genuine git fsck (parent killed, 98613 reads, well past the age threshold) and confirming it was spared while all six wedged processes were reaped.

Happy to run further instrumentation on this host if it helps; it reproduces here regularly.