fix: Do not exit MCP if the parent process is alive (#770)

closes #703

Bumped inactivity timeout to an hour and make it actually check every
minute if parent is alive and working

Cargo.lock

@@ -662,8 +662,10 @@ dependencies = [
  "schemars",
  "serde",
  "serde_json",
+ "tempfile",
  "tokio",
  "tracing",
+ "windows-sys 0.60.2",
 ]
 
 [[package]]

crates/fff-mcp/Cargo.toml

@@ -30,3 +30,13 @@ tokio = { version = "1", features = ["full"] }
 tracing = { workspace = true }
 git2 = { workspace = true }
 clap = { version = "4", features = ["derive", "env"] }
+
+[target.'cfg(windows)'.dependencies]
+windows-sys = { version = "0.60", features = [
+    "Win32_Foundation",
+    "Win32_System_Threading",
+    "Win32_System_Diagnostics_ToolHelp",
+] }
+
+[dev-dependencies]
+tempfile = "3.8"

crates/fff-mcp/src/main.rs

@@ -1,9 +1,12 @@
 mod cursor;
 mod healthcheck;
 mod output;
+mod parent;
 mod server;
 mod update_check;
 
+use std::time::{Duration, SystemTime};
+
 use clap::Parser;
 use fff::file_picker::FilePicker;
 use fff::frecency::FrecencyTracker;
@@ -92,7 +95,7 @@ pub const MCP_INSTRUCTIONS: &str = concat!(
     "  !generated/ - exclude generated code",
 );
 
-/// FFF MCP Server -- a high performance & accuracy file finder for AI code assistants.
+/// FFF MCP Server - a high performance & accuracy file finder for AI code assistants.
 #[derive(Parser)]
 #[command(name = "fff-mcp", version = concat!(env!("CARGO_PKG_VERSION"), " (", env!("FFF_GIT_HASH"), ")"))]
 pub(crate) struct Args {
@@ -183,11 +186,12 @@ pub(crate) struct Args {
     #[arg(long = "healthcheck")]
     pub(crate) healthcheck: bool,
 
-    /// Exit after this many seconds of inactivity. 0 = never exit.
+    /// Timeout of inactivity after which fff mcp will be exited. Even if the parent process
+    /// is alive we don't want to occupy resources on index and file watches if fff is unused
     #[arg(
         long = "idle-timeout-secs",
         env = "FFF_MCP_IDLE_TIMEOUT_SECS",
-        default_value_t = 900
+        default_value_t = 60 * 60
     )]
     idle_timeout_secs: u64,
 }
@@ -343,9 +347,20 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
         }
     };
 
-    if idle_timeout_secs > 0 {
+    let parent_watcher = parent::ParentWatcher::new();
+    match &parent_watcher {
+        Some(watcher) => tracing::info!(
+            "Watching parent process (pid {}); will exit when it dies",
+            watcher.parent_pid()
+        ),
+        None => tracing::warn!(
+            "Parent process liveness detection unavailable; idle timeout will exit unconditionally"
+        ),
+    }
+
+    if idle_timeout_secs > 0 || parent_watcher.is_some() {
         last_activity.store(
-            std::time::SystemTime::now()
+            SystemTime::now()
                 .duration_since(std::time::UNIX_EPOCH)
                 .map(|d| d.as_secs())
                 .unwrap_or(0),
@@ -354,9 +369,27 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
 
         let last_activity_for_watchdog = last_activity.clone();
         tokio::spawn(async move {
-            let tick = std::time::Duration::from_secs(60);
+            let tick = watchdog_interval();
             loop {
                 tokio::time::sleep(tick).await;
+
+                if let Some(ref watcher) = parent_watcher {
+                    if !watcher.parent_alive() {
+                        tracing::info!(
+                            "Parent process (pid {}) exited, shutting down",
+                            watcher.parent_pid()
+                        );
+                        flush_logs_and_exit().await;
+                    }
+                    // Parent is alive: it owns our lifecycle, never exit on idle
+                    // Clients like Codex do not restart MCP servers @see #703
+                    continue;
+                }
+
+                if idle_timeout_secs == 0 {
+                    continue;
+                }
+
                 let now = std::time::SystemTime::now()
                     .duration_since(std::time::UNIX_EPOCH)
                     .map(|d| d.as_secs())
@@ -364,12 +397,8 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
 
                 let last = last_activity_for_watchdog.load(std::sync::atomic::Ordering::Relaxed);
                 if now.saturating_sub(last) >= idle_timeout_secs {
-                    tracing::info!(
-                        "Exiting after {}s of inactivity (idle_timeout_secs={})",
-                        now.saturating_sub(last),
-                        idle_timeout_secs
-                    );
-                    std::process::exit(0);
+                    tracing::info!(?idle_timeout_secs, "Exiting due to inactivity",);
+                    flush_logs_and_exit().await;
                 }
             }
         });
@@ -396,3 +425,20 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
 
     Ok(())
 }
+
+// Tracing appender is non blocking, to get full log give it some time before hard exit
+async fn flush_logs_and_exit() -> ! {
+    tokio::time::sleep(std::time::Duration::from_millis(250)).await;
+    std::process::exit(0);
+}
+
+fn watchdog_interval() -> Duration {
+    if cfg!(debug_assertions)
+        && let Some(milliseconds) = std::env::var("FFF_MCP_TEST_WATCHDOG_INTERVAL_MS")
+            .ok()
+            .and_then(|value| value.parse().ok())
+    {
+        return Duration::from_millis(milliseconds);
+    }
+    Duration::from_secs(60)
+}

crates/fff-mcp/src/parent.rs

@@ -0,0 +1,99 @@
+#[cfg(unix)]
+mod imp {
+    pub struct ParentWatcher {
+        ppid: u32,
+    }
+
+    impl ParentWatcher {
+        pub fn new() -> Option<Self> {
+            let ppid = std::os::unix::process::parent_id();
+            // ppid <= 1 means we were spawned by init and can't detect death
+            (ppid > 1).then_some(Self { ppid })
+        }
+
+        pub fn parent_pid(&self) -> u32 {
+            self.ppid
+        }
+
+        // When the parent dies the kernel reparents us, so getppid() changes.
+        // Race-free and immune to PID reuse, unlike kill(ppid, 0).
+        pub fn parent_alive(&self) -> bool {
+            std::os::unix::process::parent_id() == self.ppid
+        }
+    }
+}
+
+#[cfg(windows)]
+mod imp {
+    use windows_sys::Win32::Foundation::{CloseHandle, HANDLE, INVALID_HANDLE_VALUE, WAIT_TIMEOUT};
+    use windows_sys::Win32::System::Diagnostics::ToolHelp::{
+        CreateToolhelp32Snapshot, PROCESSENTRY32, Process32First, Process32Next, TH32CS_SNAPPROCESS,
+    };
+    use windows_sys::Win32::System::Threading::{
+        GetCurrentProcessId, OpenProcess, PROCESS_SYNCHRONIZE, WaitForSingleObject,
+    };
+
+    pub struct ParentWatcher {
+        handle: HANDLE,
+        ppid: u32,
+    }
+
+    // HANDLE is a raw pointer; it is only ever used via WaitForSingleObject
+    // which is thread-safe, so moving/sharing the watcher across threads is fine.
+    unsafe impl Send for ParentWatcher {}
+    unsafe impl Sync for ParentWatcher {}
+
+    impl ParentWatcher {
+        pub fn new() -> Option<Self> {
+            let ppid = parent_pid_of_current()?;
+            let handle = unsafe { OpenProcess(PROCESS_SYNCHRONIZE, 0, ppid) };
+            if handle.is_null() {
+                return None;
+            }
+            // Holding the handle pins the PID, preventing reuse for the process lifetime
+            Some(Self { handle, ppid })
+        }
+
+        pub fn parent_pid(&self) -> u32 {
+            self.ppid
+        }
+
+        pub fn parent_alive(&self) -> bool {
+            unsafe { WaitForSingleObject(self.handle, 0) == WAIT_TIMEOUT }
+        }
+    }
+
+    impl Drop for ParentWatcher {
+        fn drop(&mut self) {
+            unsafe { CloseHandle(self.handle) };
+        }
+    }
+
+    fn parent_pid_of_current() -> Option<u32> {
+        unsafe {
+            let snapshot = CreateToolhelp32Snapshot(TH32CS_SNAPPROCESS, 0);
+            if snapshot == INVALID_HANDLE_VALUE {
+                return None;
+            }
+            let mut entry: PROCESSENTRY32 = std::mem::zeroed();
+            entry.dwSize = std::mem::size_of::<PROCESSENTRY32>() as u32;
+            let current = GetCurrentProcessId();
+            let mut found = None;
+            if Process32First(snapshot, &mut entry) != 0 {
+                loop {
+                    if entry.th32ProcessID == current {
+                        found = Some(entry.th32ParentProcessID);
+                        break;
+                    }
+                    if Process32Next(snapshot, &mut entry) == 0 {
+                        break;
+                    }
+                }
+            }
+            CloseHandle(snapshot);
+            found
+        }
+    }
+}
+
+pub use imp::ParentWatcher;

crates/fff-mcp/tests/parent_liveness.rs

@@ -0,0 +1,181 @@
+use std::io::{BufRead, BufReader, Write};
+use std::process::{Child, ChildStdin, Command, Stdio};
+use std::sync::mpsc;
+use std::time::{Duration, Instant};
+
+const BIN: &str = env!("CARGO_BIN_EXE_fff-mcp");
+
+#[test]
+fn stays_alive_while_parent_alive_despite_idle_timeout() {
+    let dir = tempfile::tempdir().unwrap();
+    std::fs::write(dir.path().join("a.txt"), "hello").unwrap();
+
+    let mut child = Command::new(BIN)
+        .arg(dir.path())
+        .args([
+            "--no-update-check",
+            "--no-warmup",
+            "--no-watch",
+            "--idle-timeout-secs",
+            "1",
+        ])
+        .arg("--log-file")
+        .arg(dir.path().join("test.log"))
+        .env("FFF_MCP_TEST_WATCHDOG_INTERVAL_MS", "100")
+        .stdin(Stdio::piped())
+        .stdout(Stdio::piped())
+        .stderr(Stdio::null())
+        .spawn()
+        .unwrap();
+
+    let mut stdin = child.stdin.take().unwrap();
+    let stdout_lines = spawn_line_reader(child.stdout.take().unwrap());
+    do_handshake(&mut stdin, &stdout_lines);
+
+    // Wait past the idle timeout and several watchdog ticks.
+    std::thread::sleep(Duration::from_secs(2));
+    assert!(
+        child.try_wait().unwrap().is_none(),
+        "fff-mcp exited on idle timeout even though its parent is alive"
+    );
+
+    // Closing stdin ends the transport; the server must still shut down cleanly.
+    drop(stdin);
+    wait_for_exit(&mut child, Duration::from_secs(15));
+}
+
+#[cfg(unix)]
+#[test]
+fn exits_when_parent_dies_even_without_idle_timeout() {
+    let dir = tempfile::tempdir().unwrap();
+    std::fs::write(dir.path().join("a.txt"), "hello").unwrap();
+    let log_path = dir.path().join("test.log");
+    let exit_signal = dir.path().join("exit-parent");
+
+    // Intermediary parent: sh backgrounds fff-mcp and waits until the handshake
+    // completes before dying and orphaning it.
+    let mut sh = Command::new("sh")
+        .arg("-c")
+        .arg(
+            // Preserve stdin before POSIX shells assign /dev/null to background jobs.
+            r#"exec 3<&0
+                "$1" "$2" --no-update-check --no-warmup --no-watch \
+                --idle-timeout-secs 0 --log-file "$3" <&3 &
+                while [ ! -e "$4" ]; do sleep 0.1; done"#,
+        )
+        .arg("sh")
+        .arg(BIN)
+        .arg(dir.path())
+        .arg(&log_path)
+        .arg(&exit_signal)
+        .env("FFF_MCP_TEST_WATCHDOG_INTERVAL_MS", "100")
+        .stdin(Stdio::piped())
+        .stdout(Stdio::piped())
+        .stderr(Stdio::null())
+        .spawn()
+        .unwrap();
+
+    let mut stdin = sh.stdin.take().unwrap();
+    let stdout_lines = spawn_line_reader(sh.stdout.take().unwrap());
+    do_handshake(&mut stdin, &stdout_lines);
+
+    std::fs::write(exit_signal, "").unwrap();
+    sh.wait().unwrap();
+
+    // We still hold the stdin write end, so the only exit path is the parent
+    // liveness check. EOF on stdout means fff-mcp closed it by exiting.
+    let deadline = Instant::now() + Duration::from_secs(5);
+    loop {
+        match stdout_lines.recv_timeout(deadline.saturating_duration_since(Instant::now())) {
+            Ok(_) => continue,
+            Err(mpsc::RecvTimeoutError::Disconnected) => break,
+            Err(mpsc::RecvTimeoutError::Timeout) => {
+                panic!("fff-mcp did not exit within 5s of its parent dying")
+            }
+        }
+    }
+    drop(stdin);
+
+    let logs = read_session_logs(dir.path());
+    assert!(
+        logs.contains("Parent process") && logs.contains("exited, shutting down"),
+        "expected parent-death exit reason in logs, got:\n{}",
+        logs
+    );
+}
+
+fn do_handshake(stdin: &mut ChildStdin, stdout_lines: &mpsc::Receiver<String>) {
+    let initialize = serde_json::json!({
+        "jsonrpc": "2.0",
+        "id": 1,
+        "method": "initialize",
+        "params": {
+            "protocolVersion": "2024-11-05",
+            "capabilities": {},
+            "clientInfo": { "name": "parent-liveness-test", "version": "0.0.0" }
+        }
+    });
+    writeln!(stdin, "{}", initialize).unwrap();
+    stdin.flush().unwrap();
+
+    let response = stdout_lines
+        .recv_timeout(Duration::from_secs(30))
+        .expect("no initialize response within 30s");
+    assert!(
+        response.contains("\"serverInfo\""),
+        "unexpected initialize response: {}",
+        response
+    );
+
+    writeln!(
+        stdin,
+        "{}",
+        serde_json::json!({ "jsonrpc": "2.0", "method": "notifications/initialized" })
+    )
+    .unwrap();
+    stdin.flush().unwrap();
+}
+
+fn spawn_line_reader(stdout: std::process::ChildStdout) -> mpsc::Receiver<String> {
+    let (tx, rx) = mpsc::channel();
+    std::thread::spawn(move || {
+        for line in BufReader::new(stdout).lines() {
+            match line {
+                Ok(line) => {
+                    if tx.send(line).is_err() {
+                        break;
+                    }
+                }
+                Err(_) => break,
+            }
+        }
+    });
+    rx
+}
+
+fn wait_for_exit(child: &mut Child, timeout: Duration) {
+    let deadline = Instant::now() + timeout;
+    while Instant::now() < deadline {
+        if child.try_wait().unwrap().is_some() {
+            return;
+        }
+        std::thread::sleep(Duration::from_millis(100));
+    }
+    child.kill().ok();
+    panic!(
+        "fff-mcp did not exit within {:?} after stdin closed",
+        timeout
+    );
+}
+
+#[cfg(unix)]
+fn read_session_logs(dir: &std::path::Path) -> String {
+    let mut combined = String::new();
+    for entry in std::fs::read_dir(dir).unwrap().flatten() {
+        let name = entry.file_name().to_string_lossy().to_string();
+        if name.starts_with("test") && name.ends_with(".log") {
+            combined.push_str(&std::fs::read_to_string(entry.path()).unwrap_or_default());
+        }
+    }
+    combined
+}