util: Use job objects to reap spawned process trees on Windows (#58885)

On Windows, `util::process::Child::kill()` only terminated the direct
child process, and nothing tied the lifetime of spawned process trees to
Zed. External agent servers launched through ACP (e.g. `claude-code-acp`
via `npx`) spawn node workers and MCP servers as grandchildren, which
were orphaned on every session teardown and accumulated indefinitely —
hundreds of idle `node.exe` processes and GBs of RAM over days.

This PR assigns spawned processes to a Win32 job object configured with
`JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE`:

- `kill()` now calls `TerminateJobObject`, killing the entire tree
instead of just the (shell wrapper) child.
- Dropping `Child` closes the job handle, which makes the OS reap the
tree — including when Zed exits for any reason (even crashes), since the
OS closes its handles.
- Unix behavior (process groups via `killpg`) is unchanged.

Added two Windows tests that spawn a real `powershell -> ping` process
tree and assert the grandchild is terminated on `kill()` and on drop.
Both fail without the fix and pass with it. Verified with `cargo test -p
util`, `script/clippy -p util`, and `cargo check -p agent_servers -p
dap` on Windows 11.

Closes #58873

Release Notes:

- Fixed external agent servers and debug adapters leaking helper
processes (e.g. node workers and MCP servers) on Windows.

---------

Co-authored-by: zed-zippy[bot] <234243425+zed-zippy[bot]@users.noreply.github.com>
Co-authored-by: Smit Barmase <heysmitbarmase@gmail.com>

Cargo.lock

@@ -19795,6 +19795,7 @@ dependencies = [
  "util_macros",
  "walkdir",
  "which 6.0.3",
+ "windows 0.61.3",
 ]
 
 [[package]]

Cargo.toml

@@ -869,6 +869,7 @@ features = [
     "Win32_System_Diagnostics_Debug",
     "Win32_System_DataExchange",
     "Win32_System_IO",
+    "Win32_System_JobObjects",
     "Win32_System_LibraryLoader",
     "Win32_System_Memory",
     "Win32_System_Ole",

crates/dap/src/transport.rs

@@ -617,8 +617,8 @@ impl Transport for TcpTransport {
                                 if has_process {
                                     let status = process.lock().as_mut().unwrap().try_status();
                                     if let Ok(Some(_)) = status {
-                                        let process = process.lock().take().unwrap().into_inner();
-                                        let output = process.output().await?;
+                                        let child = process.lock().take().unwrap();
+                                        let output = child.output().await?;
                                         let output = if output.stderr.is_empty() {
                                             String::from_utf8_lossy(&output.stdout).to_string()
                                         } else {

crates/util/Cargo.toml

@@ -60,6 +60,7 @@ mach2.workspace = true
 
 [target.'cfg(windows)'.dependencies]
 tendril = "0.4.3"
+windows.workspace = true
 
 [dev-dependencies]
 rand.workspace = true

crates/util/src/process.rs

@@ -2,9 +2,17 @@ use anyhow::{Context as _, Result};
 use std::process::Stdio;
 
 /// A wrapper around `smol::process::Child` that ensures all subprocesses
-/// are killed when the process is terminated by using process groups.
+/// are killed when the process is terminated: on Unix by using process
+/// groups, and on Windows by using job objects.
+///
+/// On Windows, dropping this struct closes the job object handle, which
+/// terminates all processes in the job. This also applies when the Zed
+/// process exits for any reason (including crashes), since the OS closes
+/// its handles, so spawned process trees can never outlive Zed.
 pub struct Child {
     process: smol::process::Child,
+    #[cfg(windows)]
+    job: Option<windows_job::JobObject>,
 }
 
 impl std::ops::Deref for Child {
@@ -52,8 +60,6 @@ impl Child {
         stdout: Stdio,
         stderr: Stdio,
     ) -> Result<Self> {
-        // TODO(windows): create a job object and add the child process handle to it,
-        // see https://learn.microsoft.com/en-us/windows/win32/procthread/job-objects
         let mut command = smol::process::Command::from(command);
         let process = command
             .stdin(stdin)
@@ -67,11 +73,42 @@ impl Child {
                 )
             })?;
 
-        Ok(Self { process })
+        // Assign the child to a job object configured to kill the entire
+        // process tree when the last job handle is closed, so descendants
+        // (e.g. node workers and MCP servers spawned by agent servers) are
+        // reaped even if the direct child doesn't clean them up. Any process
+        // the child spawns after this assignment is automatically part of the
+        // job.
+        //
+        // There is a small race: descendants the child spawns between the
+        // `spawn()` call returning and the assignment below escape the job.
+        // Closing it fully would require creating the process suspended
+        // (`CREATE_SUSPENDED`), assigning it, then resuming it, which the
+        // std/smol process APIs don't support without reimplementing process
+        // creation. The window is microseconds, and the children we care
+        // about (`npx`, `node`, etc.) take far longer to load their runtime
+        // and spawn anything, so in practice nothing escapes.
+        let job = windows_job::JobObject::new()
+            .and_then(|job| {
+                job.assign_process(process.id())?;
+                Ok(job)
+            })
+            .map_err(|error| {
+                log::error!("failed to assign spawned process to a job object: {error:#}");
+            })
+            .ok();
+
+        Ok(Self { process, job })
     }
 
-    pub fn into_inner(self) -> smol::process::Child {
-        self.process
+    /// Consumes the child, draining its stdout/stderr and waiting for it to
+    /// exit, then returns the collected output.
+    pub async fn output(self) -> Result<std::process::Output> {
+        // NOTE: Keep `self` alive across this await, do not destructure it to
+        // pull `process` out first. On Windows that drops the job object early,
+        // which triggers `JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE` and kills the
+        // child before `output()` finishes collecting its stdout/stderr.
+        Ok(self.process.output().await?)
     }
 
     #[cfg(not(windows))]
@@ -85,8 +122,175 @@ impl Child {
 
     #[cfg(windows)]
     pub fn kill(&mut self) -> Result<()> {
-        // TODO(windows): terminate the job object in kill
-        self.process.kill()?;
-        Ok(())
+        if let Some(job) = &self.job {
+            job.terminate()
+        } else {
+            self.process.kill()?;
+            Ok(())
+        }
+    }
+}
+
+#[cfg(windows)]
+mod windows_job {
+    use crate::ResultExt as _;
+    use anyhow::{Context as _, Result};
+    use windows::Win32::{
+        Foundation::{CloseHandle, HANDLE},
+        System::{
+            JobObjects::{
+                AssignProcessToJobObject, CreateJobObjectW, JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE,
+                JOBOBJECT_EXTENDED_LIMIT_INFORMATION, JobObjectExtendedLimitInformation,
+                SetInformationJobObject, TerminateJobObject,
+            },
+            Threading::{OpenProcess, PROCESS_SET_QUOTA, PROCESS_TERMINATE},
+        },
+    };
+
+    /// A Win32 job object configured with `JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE`:
+    /// all processes assigned to the job (and their descendants) are terminated
+    /// when the last handle to the job is closed, which happens when this struct
+    /// is dropped, or when the OS closes the owning process's handles after it
+    /// exits for any reason.
+    pub(crate) struct JobObject(HANDLE);
+
+    // SAFETY: Job object handles can be used from any thread.
+    unsafe impl Send for JobObject {}
+    unsafe impl Sync for JobObject {}
+
+    impl JobObject {
+        pub(crate) fn new() -> Result<Self> {
+            unsafe {
+                let job =
+                    Self(CreateJobObjectW(None, None).context("failed to create job object")?);
+                let mut info = JOBOBJECT_EXTENDED_LIMIT_INFORMATION::default();
+                info.BasicLimitInformation.LimitFlags = JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE;
+                SetInformationJobObject(
+                    job.0,
+                    JobObjectExtendedLimitInformation,
+                    &info as *const _ as *const _,
+                    size_of::<JOBOBJECT_EXTENDED_LIMIT_INFORMATION>() as u32,
+                )
+                .context("failed to set job object limits")?;
+                Ok(job)
+            }
+        }
+
+        pub(crate) fn assign_process(&self, pid: u32) -> Result<()> {
+            unsafe {
+                let process = OpenProcess(PROCESS_SET_QUOTA | PROCESS_TERMINATE, false, pid)
+                    .context("failed to open process")?;
+                let result = AssignProcessToJobObject(self.0, process)
+                    .context("failed to assign process to job object");
+                CloseHandle(process).log_err();
+                result
+            }
+        }
+
+        pub(crate) fn terminate(&self) -> Result<()> {
+            unsafe { TerminateJobObject(self.0, 1).context("failed to terminate job object") }
+        }
+    }
+
+    impl Drop for JobObject {
+        fn drop(&mut self) {
+            unsafe {
+                CloseHandle(self.0).log_err();
+            }
+        }
+    }
+}
+
+#[cfg(all(test, windows))]
+mod windows_tests {
+    use super::*;
+    use std::time::{Duration, Instant};
+
+    /// Spawns a process tree `powershell -> ping` via `Child::spawn` and
+    /// returns the `Child` along with the pid of the grandchild (`ping`).
+    fn spawn_process_tree(temp_dir: &std::path::Path) -> (Child, u32) {
+        let pid_file = temp_dir.join("grandchild_pid");
+        let mut command = std::process::Command::new("powershell.exe");
+        command.args(["-NoProfile", "-Command"]).arg(format!(
+            "$p = Start-Process -FilePath ping.exe -ArgumentList @('-n','60','127.0.0.1') -PassThru -WindowStyle Hidden; \
+             Set-Content -LiteralPath '{}' -Value $p.Id; \
+             Wait-Process -Id $p.Id",
+            pid_file.display()
+        ));
+        let child = Child::spawn(command, Stdio::null(), Stdio::null(), Stdio::null())
+            .expect("failed to spawn powershell");
+
+        let deadline = Instant::now() + Duration::from_secs(5);
+        let grandchild_pid = loop {
+            if let Ok(contents) = std::fs::read_to_string(&pid_file)
+                && let Ok(pid) = contents.trim().parse::<u32>()
+            {
+                break pid;
+            }
+            assert!(
+                Instant::now() < deadline,
+                "timed out waiting for grandchild pid file"
+            );
+            std::thread::sleep(Duration::from_millis(50));
+        };
+        assert!(
+            process_is_alive(grandchild_pid),
+            "grandchild should be alive after spawning"
+        );
+        (child, grandchild_pid)
+    }
+
+    fn process_is_alive(pid: u32) -> bool {
+        use windows::Win32::{
+            Foundation::{CloseHandle, STILL_ACTIVE},
+            System::Threading::{
+                GetExitCodeProcess, OpenProcess, PROCESS_QUERY_LIMITED_INFORMATION,
+            },
+        };
+
+        unsafe {
+            let Ok(handle) = OpenProcess(PROCESS_QUERY_LIMITED_INFORMATION, false, pid) else {
+                return false;
+            };
+            let mut exit_code = 0u32;
+            let alive = GetExitCodeProcess(handle, &mut exit_code).is_ok()
+                && exit_code == STILL_ACTIVE.0 as u32;
+            CloseHandle(handle).expect("failed to close process handle");
+            alive
+        }
+    }
+
+    fn assert_process_exits(pid: u32, message: &str) {
+        let deadline = Instant::now() + Duration::from_secs(2);
+        while process_is_alive(pid) {
+            assert!(Instant::now() < deadline, "{message} (pid {pid})");
+            std::thread::sleep(Duration::from_millis(100));
+        }
+    }
+
+    #[test]
+    fn test_kill_terminates_grandchildren() {
+        let temp_dir = tempfile::tempdir().unwrap();
+        let (mut child, grandchild_pid) = spawn_process_tree(temp_dir.path());
+
+        child.kill().expect("failed to kill child");
+
+        assert_process_exits(
+            grandchild_pid,
+            "grandchild should be terminated after killing the child",
+        );
+    }
+
+    #[test]
+    fn test_drop_terminates_grandchildren() {
+        let temp_dir = tempfile::tempdir().unwrap();
+        let (child, grandchild_pid) = spawn_process_tree(temp_dir.path());
+
+        drop(child);
+
+        assert_process_exits(
+            grandchild_pid,
+            "grandchild should be terminated after dropping the child",
+        );
     }
 }