Files
codex/codex-rs/app-server/src/fs_watch.rs
T
Eric TrautandGitHub 522f549922 Fix fs/watch debounce batching (#24716)
## Summary

`fs/watch` was using a local debounce wrapper whose deadline was
initialized once and then reused after the first batch. Once that stale
deadline was in the past, later file changes could bypass the intended
200ms debounce and send noisier `fs/changed` notifications.

This moves the debounce wrapper into `codex-file-watcher` as
`DebouncedWatchReceiver`, resets the debounce deadline for each event
batch, preserves pending paths across cancelled receives, and updates
app-server `fs/watch` to use the shared wrapper.

Fixes #24692.
2026-05-28 23:09:55 -07:00

378 lines
12 KiB
Rust

use crate::error_code::invalid_request;
use crate::outgoing_message::ConnectionId;
use crate::outgoing_message::OutgoingMessageSender;
use codex_app_server_protocol::FsChangedNotification;
use codex_app_server_protocol::FsUnwatchParams;
use codex_app_server_protocol::FsUnwatchResponse;
use codex_app_server_protocol::FsWatchParams;
use codex_app_server_protocol::FsWatchResponse;
use codex_app_server_protocol::JSONRPCErrorError;
use codex_app_server_protocol::ServerNotification;
use codex_file_watcher::DebouncedWatchReceiver;
use codex_file_watcher::FileWatcher;
use codex_file_watcher::FileWatcherSubscriber;
use codex_file_watcher::WatchPath;
use codex_file_watcher::WatchRegistration;
use std::collections::HashMap;
use std::collections::hash_map::Entry;
use std::hash::Hash;
use std::sync::Arc;
use std::time::Duration;
use tokio::sync::Mutex as AsyncMutex;
#[cfg(test)]
use tokio::sync::mpsc;
use tokio::sync::oneshot;
use tracing::warn;
const FS_CHANGED_NOTIFICATION_DEBOUNCE: Duration = Duration::from_millis(200);
#[derive(Clone)]
pub(crate) struct FsWatchManager {
outgoing: Arc<OutgoingMessageSender>,
file_watcher: Arc<FileWatcher>,
state: Arc<AsyncMutex<FsWatchState>>,
}
#[derive(Default)]
struct FsWatchState {
entries: HashMap<WatchKey, WatchEntry>,
}
struct WatchEntry {
terminate_tx: oneshot::Sender<oneshot::Sender<()>>,
_subscriber: FileWatcherSubscriber,
_registration: WatchRegistration,
}
#[derive(Clone, Debug, Eq, Hash, PartialEq)]
struct WatchKey {
connection_id: ConnectionId,
watch_id: String,
}
impl FsWatchManager {
pub(crate) fn new(outgoing: Arc<OutgoingMessageSender>) -> Self {
let file_watcher = match FileWatcher::new() {
Ok(file_watcher) => Arc::new(file_watcher),
Err(err) => {
warn!("filesystem watch manager falling back to noop core watcher: {err}");
Arc::new(FileWatcher::noop())
}
};
Self::new_with_file_watcher(outgoing, file_watcher)
}
fn new_with_file_watcher(
outgoing: Arc<OutgoingMessageSender>,
file_watcher: Arc<FileWatcher>,
) -> Self {
Self {
outgoing,
file_watcher,
state: Arc::new(AsyncMutex::new(FsWatchState::default())),
}
}
pub(crate) async fn watch(
&self,
connection_id: ConnectionId,
params: FsWatchParams,
) -> Result<FsWatchResponse, JSONRPCErrorError> {
let watch_id = params.watch_id;
let watch_key = WatchKey {
connection_id,
watch_id: watch_id.clone(),
};
let outgoing = self.outgoing.clone();
let (subscriber, rx) = self.file_watcher.add_subscriber();
let watch_root = params.path.clone();
let registration = subscriber.register_paths(vec![WatchPath {
path: params.path.to_path_buf(),
recursive: false,
}]);
let (terminate_tx, terminate_rx) = oneshot::channel();
match self.state.lock().await.entries.entry(watch_key) {
Entry::Occupied(_) => {
return Err(invalid_request(format!(
"watchId already exists: {watch_id}"
)));
}
Entry::Vacant(entry) => {
entry.insert(WatchEntry {
terminate_tx,
_subscriber: subscriber,
_registration: registration,
});
}
}
let task_watch_id = watch_id.clone();
tokio::spawn(async move {
let mut rx = DebouncedWatchReceiver::new(rx, FS_CHANGED_NOTIFICATION_DEBOUNCE);
tokio::pin!(terminate_rx);
loop {
let event = tokio::select! {
biased;
_ = &mut terminate_rx => break,
event = rx.recv() => match event {
Some(event) => event,
None => break,
},
};
let mut changed_paths = event
.paths
.into_iter()
.map(|path| watch_root.join(path))
.collect::<Vec<_>>();
changed_paths.sort_by(|left, right| left.as_path().cmp(right.as_path()));
if !changed_paths.is_empty() {
outgoing
.send_server_notification_to_connection_and_wait(
connection_id,
ServerNotification::FsChanged(FsChangedNotification {
watch_id: task_watch_id.clone(),
changed_paths,
}),
)
.await;
}
}
});
Ok(FsWatchResponse { path: params.path })
}
pub(crate) async fn unwatch(
&self,
connection_id: ConnectionId,
params: FsUnwatchParams,
) -> Result<FsUnwatchResponse, JSONRPCErrorError> {
let watch_key = WatchKey {
connection_id,
watch_id: params.watch_id,
};
let entry = self.state.lock().await.entries.remove(&watch_key);
if let Some(entry) = entry {
// Wait for the oneshot to be destroyed by the task to ensure that no notifications
// are send after the unwatch response.
let (done_tx, done_rx) = oneshot::channel();
let _ = entry.terminate_tx.send(done_tx);
let _ = done_rx.await;
}
Ok(FsUnwatchResponse {})
}
pub(crate) async fn connection_closed(&self, connection_id: ConnectionId) {
let mut state = self.state.lock().await;
state
.entries
.extract_if(|key, _| key.connection_id == connection_id)
.count();
}
}
#[cfg(test)]
mod tests {
use super::*;
use codex_utils_absolute_path::AbsolutePathBuf;
use pretty_assertions::assert_eq;
use std::collections::HashSet;
use std::path::PathBuf;
use tempfile::TempDir;
fn absolute_path(path: PathBuf) -> AbsolutePathBuf {
assert!(
path.is_absolute(),
"path must be absolute: {}",
path.display()
);
AbsolutePathBuf::try_from(path).expect("path should be absolute")
}
fn manager_with_noop_watcher() -> FsWatchManager {
const OUTGOING_BUFFER: usize = 1;
let (tx, _rx) = mpsc::channel(OUTGOING_BUFFER);
FsWatchManager::new_with_file_watcher(
Arc::new(OutgoingMessageSender::new(
tx,
codex_analytics::AnalyticsEventsClient::disabled(),
)),
Arc::new(FileWatcher::noop()),
)
}
#[tokio::test]
async fn watch_uses_client_id_and_tracks_the_owner_scoped_entry() {
let temp_dir = TempDir::new().expect("temp dir");
let head_path = temp_dir.path().join("HEAD");
std::fs::write(&head_path, "ref: refs/heads/main\n").expect("write HEAD");
let manager = manager_with_noop_watcher();
let path = absolute_path(head_path);
let watch_id = "watch-head".to_string();
let response = manager
.watch(
ConnectionId(1),
FsWatchParams {
watch_id: watch_id.clone(),
path: path.clone(),
},
)
.await
.expect("watch should succeed");
assert_eq!(response.path, path);
let state = manager.state.lock().await;
assert_eq!(
state.entries.keys().cloned().collect::<HashSet<_>>(),
HashSet::from([WatchKey {
connection_id: ConnectionId(1),
watch_id,
}])
);
}
#[tokio::test]
async fn unwatch_is_scoped_to_the_connection_that_created_the_watch() {
let temp_dir = TempDir::new().expect("temp dir");
let head_path = temp_dir.path().join("HEAD");
std::fs::write(&head_path, "ref: refs/heads/main\n").expect("write HEAD");
let manager = manager_with_noop_watcher();
manager
.watch(
ConnectionId(1),
FsWatchParams {
watch_id: "watch-head".to_string(),
path: absolute_path(head_path),
},
)
.await
.expect("watch should succeed");
let watch_key = WatchKey {
connection_id: ConnectionId(1),
watch_id: "watch-head".to_string(),
};
manager
.unwatch(
ConnectionId(2),
FsUnwatchParams {
watch_id: "watch-head".to_string(),
},
)
.await
.expect("foreign unwatch should be a no-op");
assert!(manager.state.lock().await.entries.contains_key(&watch_key));
manager
.unwatch(
ConnectionId(1),
FsUnwatchParams {
watch_id: "watch-head".to_string(),
},
)
.await
.expect("owner unwatch should succeed");
assert!(!manager.state.lock().await.entries.contains_key(&watch_key));
}
#[tokio::test]
async fn watch_rejects_duplicate_id_for_the_same_connection() {
let temp_dir = TempDir::new().expect("temp dir");
let head_path = temp_dir.path().join("HEAD");
let fetch_head_path = temp_dir.path().join("FETCH_HEAD");
std::fs::write(&head_path, "ref: refs/heads/main\n").expect("write HEAD");
std::fs::write(&fetch_head_path, "old-fetch\n").expect("write FETCH_HEAD");
let manager = manager_with_noop_watcher();
manager
.watch(
ConnectionId(1),
FsWatchParams {
watch_id: "watch-head".to_string(),
path: absolute_path(head_path),
},
)
.await
.expect("first watch should succeed");
let error = manager
.watch(
ConnectionId(1),
FsWatchParams {
watch_id: "watch-head".to_string(),
path: absolute_path(fetch_head_path),
},
)
.await
.expect_err("duplicate watch should fail");
assert_eq!(error.message, "watchId already exists: watch-head");
assert_eq!(manager.state.lock().await.entries.len(), 1);
}
#[tokio::test]
async fn connection_closed_removes_only_that_connections_watches() {
let temp_dir = TempDir::new().expect("temp dir");
let head_path = temp_dir.path().join("HEAD");
let fetch_head_path = temp_dir.path().join("FETCH_HEAD");
let packed_refs_path = temp_dir.path().join("packed-refs");
std::fs::write(&head_path, "ref: refs/heads/main\n").expect("write HEAD");
std::fs::write(&fetch_head_path, "old-fetch\n").expect("write FETCH_HEAD");
std::fs::write(&packed_refs_path, "refs\n").expect("write packed-refs");
let manager = manager_with_noop_watcher();
let response = manager
.watch(
ConnectionId(1),
FsWatchParams {
watch_id: "watch-head".to_string(),
path: absolute_path(head_path.clone()),
},
)
.await
.expect("first watch should succeed");
manager
.watch(
ConnectionId(1),
FsWatchParams {
watch_id: "watch-fetch-head".to_string(),
path: absolute_path(fetch_head_path),
},
)
.await
.expect("second watch should succeed");
manager
.watch(
ConnectionId(2),
FsWatchParams {
watch_id: "watch-packed-refs".to_string(),
path: absolute_path(packed_refs_path),
},
)
.await
.expect("third watch should succeed");
manager.connection_closed(ConnectionId(1)).await;
assert_eq!(
manager
.state
.lock()
.await
.entries
.keys()
.cloned()
.collect::<HashSet<_>>(),
HashSet::from([WatchKey {
connection_id: ConnectionId(2),
watch_id: "watch-packed-refs".to_string(),
}])
);
assert_eq!(response.path, absolute_path(head_path));
}
}