From b36c0b11896b0708035b0778de5635b45e84fe0d Mon Sep 17 00:00:00 2001
From: Teddy Ni <tni@openai.com>
Date: Mon, 13 Jul 2026 17:28:26 +0000
Subject: [PATCH] Reap exited PID-managed app-server children (#32838)

## Why

An app-server child started by a previous updater can remain as a zombie after
termination. Its PID then still appears active, causing daemon shutdown to wait
until its timeout instead of removing the stale PID file.

## What changed

- Poll `waitpid` with `WNOHANG` while waiting for a PID-managed process to stop.
- Check process state before enforcing the stop deadline so an exit at the
  deadline can still be observed and cleaned up.

## Testing

Added a regression test that stops an untracked child and verifies that shutdown
finishes promptly and removes its PID file.

GitOrigin-RevId: b203cf860e4207716a6cfc36015f8ee77918d13a
---
 codex-rs/app-server-daemon/src/backend/pid.rs | 12 +++++-
 .../src/backend/pid_tests.rs                  | 41 +++++++++++++++++++
 2 files changed, 52 insertions(+), 1 deletion(-)

diff --git a/codex-rs/app-server-daemon/src/backend/pid.rs b/codex-rs/app-server-daemon/src/backend/pid.rs
index b525e61254..aa05e1db8e 100644
--- a/codex-rs/app-server-daemon/src/backend/pid.rs
+++ b/codex-rs/app-server-daemon/src/backend/pid.rs
@@ -254,13 +254,23 @@ impl PidBackend {
             let started_at = tokio::time::Instant::now();
             let deadline = tokio::time::Instant::now() + STOP_TIMEOUT;
             let mut forced = false;
-            while tokio::time::Instant::now() < deadline {
+            loop {
+                #[cfg(unix)]
+                if let Ok(raw_pid) = libc::pid_t::try_from(pid)
+                    && raw_pid > 0
+                {
+                    // A previous updater may have started this child; reap it if it has exited.
+                    unsafe { libc::waitpid(raw_pid, std::ptr::null_mut(), libc::WNOHANG) };
+                }
                 if !self.record_is_active(&record).await? {
                     match self.refresh_after_stale_record(&record).await? {
                         PidFileState::Missing => return Ok(()),
                         PidFileState::Starting | PidFileState::Running(_) => break,
                     }
                 }
+                if tokio::time::Instant::now() >= deadline {
+                    break;
+                }
                 if !forced && started_at.elapsed() >= STOP_GRACE_PERIOD {
                     self.force_terminate_process(pid)?;
                     forced = true;
diff --git a/codex-rs/app-server-daemon/src/backend/pid_tests.rs b/codex-rs/app-server-daemon/src/backend/pid_tests.rs
index 57b3d227e9..de67fa4e3d 100644
--- a/codex-rs/app-server-daemon/src/backend/pid_tests.rs
+++ b/codex-rs/app-server-daemon/src/backend/pid_tests.rs
@@ -1,3 +1,4 @@
+use std::process::Stdio;
 use std::time::Duration;
 
 use pretty_assertions::assert_eq;
@@ -10,6 +11,7 @@ use super::PidCommandKind;
 use super::PidFileState;
 use super::PidLogTail;
 use super::PidRecord;
+use super::read_process_start_time;
 use super::read_stderr_log_tail;
 use super::stderr_log_file_for_pid_file;
 use super::try_lock_file;
@@ -147,6 +149,45 @@ async fn stale_record_cleanup_preserves_replacement_record() {
     );
 }
 
+#[tokio::test]
+async fn stop_reaps_untracked_app_server_child() {
+    let temp_dir = TempDir::new().expect("temp dir");
+    let pid_file = temp_dir.path().join("app-server.pid");
+    let mut child = std::process::Command::new("sleep")
+        .arg("5")
+        .stdin(Stdio::null())
+        .stdout(Stdio::null())
+        .stderr(Stdio::null())
+        .spawn()
+        .expect("spawn app-server shim");
+    let pid = child.id();
+    let record = PidRecord {
+        pid,
+        process_start_time: read_process_start_time(pid).await.expect("start time"),
+    };
+    tokio::fs::write(
+        &pid_file,
+        serde_json::to_vec(&record).expect("serialize pid"),
+    )
+    .await
+    .expect("write pid file");
+    let backend = PidBackend::new(
+        temp_dir.path().join("codex"),
+        pid_file.clone(),
+        /*remote_control_enabled*/ false,
+    );
+
+    let result = tokio::time::timeout(Duration::from_secs(2), backend.stop()).await;
+    if matches!(child.try_wait(), Ok(None)) {
+        let _ = child.kill();
+        let _ = child.wait();
+    }
+
+    // `sleep` is not tracked by Tokio, so stop must reap it instead of leaving a zombie.
+    result.expect("stop timed out").expect("stop");
+    assert!(!pid_file.exists());
+}
+
 #[test]
 fn update_loop_uses_hidden_app_server_subcommand() {
     let backend = PidBackend {