slipcase_open/

watch.rs

//! Watching the content directory, and what the events in it mean.
//
// Author: David M. Anderson
// Built with AI assistance (Claude, Anthropic)
//
//! **The watch is on the directory and never on the file**, which concept 6
//! calls one of the three things that make write-back detection hard. A serious
//! editor saves by writing a temporary sibling and renaming it over the target,
//! so a watcher registered on the content file loses its handle on the first
//! save and never fires again. `notify` will watch a directory on all three
//! platforms, but only if it is asked to.
//!
//! ## The sibling signal
//!
//! Concept 6.1: the content directory holds one file, put there by this tool,
//! so anything else appearing in it was created by the target application — a
//! lock file, an autosave, a backup, a save in progress. Nothing here needs to
//! know which, or what any of them are called, which is why there is no table
//! of `~$name.docx` and `.~lock.name#` conventions to maintain and no
//! application it fails to know about.
//!
//! Siblings present means the application is working in there, which process
//! exit does not tell you. Siblings gone means it has cleaned up and has
//! probably finished. It stays a heuristic in both directions: most
//! read-oriented applications write no sibling at all, so an empty directory
//! means nothing, and an application that leaves a backup behind for good never
//! produces the cleaned-up signal. Both degrade to silence, which is the
//! intended fallback and is why the session model does not rest on this.

use std::path::Path;
use std::sync::mpsc::{self, Receiver};
use std::time::Duration;

use notify::{RecommendedWatcher, RecursiveMode, Watcher as _};

/// What happened in a content directory.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Change {
    /// The content file itself was written, replaced, or removed. The
    /// write-back trigger.
    Content,
    /// Something the target application made appeared beside it.
    SiblingAppeared,
    /// Something it had made went away.
    SiblingWentAway,
}

/// What an event in the content directory means, given the content file's name.
///
/// Pure, so the rule is testable without a filesystem or a race. Paths are
/// compared by their final component: `notify` reports absolute paths, and a
/// rename within the directory arrives as paths that differ only there.
///
/// A rename over the content file produces events naming both the temporary
/// sibling and the content file, and both are worth reporting — the first says
/// the application is working, the second is the save.
#[must_use]
pub fn classify(content_name: &str, paths: &[&Path], kind: EventKind) -> Vec<Change> {
    paths
        .iter()
        .map(|p| {
            let is_content = p.file_name().is_some_and(|n| n == content_name);
            match (is_content, kind) {
                (true, _) => Change::Content,
                (false, EventKind::Gone) => Change::SiblingWentAway,
                (false, _) => Change::SiblingAppeared,
            }
        })
        .collect()
}

/// The shape of an event, reduced to what the rule above needs.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum EventKind {
    /// Created, written, or renamed into place.
    Touched,
    /// Removed, or renamed away.
    Gone,
}

impl EventKind {
    /// Reduce one of `notify`'s events, or discard it.
    ///
    /// **Reading a file is not changing it.** A plain read of the watched
    /// content file emits `Access(Open(Any))` on Linux — measured on
    /// 2026-08-30 — and treating that as a save makes every reader of the
    /// content file a source of spurious write-backs. Anything that reads it
    /// counts: the write-back itself opens the content file, and so does
    /// `recover::state`, which is called once per session by `sessions`.
    /// Running `sessions` in a loop beside an open session produced a repack
    /// per poll before this arm existed.
    ///
    /// The repacks are invisible from outside, which is why this is a rule and
    /// a test rather than something anyone would notice: each one writes the
    /// same bytes, so the container stays right and only the work is wrong.
    ///
    /// The exception is a close after writing, which is the one access event
    /// that means a save finished. Linux reports it and the other platforms do
    /// not, so it is a signal to take where it is offered rather than one to
    /// depend on.
    ///
    /// `Any` and `Other` stay a touch. An unrecognised event in this directory
    /// is still something happening in it, and the cost of treating one as a
    /// save is a repack that writes what is already there.
    fn of(kind: notify::EventKind) -> Option<Self> {
        use notify::event::{AccessKind, AccessMode, ModifyKind, RenameMode};
        match kind {
            // A removal, and the half of a rename that names where the file
            // was. The other half names where it went, which is a touch.
            notify::EventKind::Remove(_)
            | notify::EventKind::Modify(ModifyKind::Name(RenameMode::From)) => Some(Self::Gone),
            notify::EventKind::Access(AccessKind::Close(AccessMode::Write)) => Some(Self::Touched),
            notify::EventKind::Access(_) => None,
            _ => Some(Self::Touched),
        }
    }
}

/// A watch on one content directory.
///
/// Holds the platform watcher, which stops when this is dropped.
pub struct Watch {
    /// `Option` so [`Drop`] can drop the watcher and *then* wait for the
    /// platform to finish stopping it.
    watcher: Option<RecommendedWatcher>,
    changes: Receiver<Change>,
    /// Where Windows reports that a watch has finished stopping.
    ///
    /// `notify`'s own constructor throws this receiver away, which is why the
    /// stop is unobservable through the ordinary API and why this builds the
    /// watcher the long way round. [`Drop`] says what it is for.
    #[cfg(windows)]
    stopped: Receiver<notify::windows::MetaEvent>,
}

/// How long [`Watch::drop`] will wait for a stop to finish.
///
/// A stop that has gone wrong should not become a hang of its own, which is
/// the defect this is here to avoid rather than to relocate. A stop that is
/// working takes microseconds, so a wait this long is only ever paid by one
/// that is not.
#[cfg(windows)]
const STOP_WAIT: Duration = Duration::from_secs(5);

impl Watch {
    /// Watch `dir` for changes to `content_name` and to anything beside it.
    ///
    /// Non-recursive: the content directory has no subdirectories of this
    /// tool's making, and an application that creates one has still created a
    /// sibling, which is the signal either way.
    ///
    /// # Errors
    ///
    /// Where the platform watcher cannot be created or cannot watch `dir`.
    pub fn on(dir: &Path, content_name: &str) -> notify::Result<Self> {
        let (tx, changes) = mpsc::channel();
        let content_name = content_name.to_string();
        let handler = move |event: notify::Result<notify::Event>| {
            let Ok(event) = event else {
                // A dropped or errored event is not a reason to tear down the
                // watch. Concept 6.2 exists because detection is unreliable,
                // and the session close is the backstop for everything this
                // misses.
                return;
            };
            let Some(kind) = EventKind::of(event.kind) else {
                return;
            };
            let paths: Vec<&Path> = event.paths.iter().map(AsRef::as_ref).collect();
            for change in classify(&content_name, &paths, kind) {
                // A closed receiver means the session is gone and there is
                // nobody to tell.
                if tx.send(change).is_err() {
                    return;
                }
            }
        };

        // On Windows the watcher is built through `create` rather than
        // `recommended_watcher`, for the one thing `recommended_watcher` throws
        // away: the channel the backend reports `SingleWatchComplete` on.
        // Without it a stop cannot be waited for, and `Drop` has to be able to
        // wait. Everywhere else the ordinary constructor is right.
        #[cfg(windows)]
        let (mut watcher, stopped) = {
            let (meta_tx, stopped) = mpsc::channel();
            let handler: std::sync::Arc<std::sync::Mutex<dyn notify::EventHandler>> =
                std::sync::Arc::new(std::sync::Mutex::new(handler));
            let watcher = notify::windows::ReadDirectoryChangesWatcher::create(handler, meta_tx)?;
            (watcher, stopped)
        };
        #[cfg(not(windows))]
        let mut watcher = notify::recommended_watcher(handler)?;

        watcher.watch(dir, RecursiveMode::NonRecursive)?;
        Ok(Self {
            watcher: Some(watcher),
            changes,
            #[cfg(windows)]
            stopped,
        })
    }

    /// Every change that has arrived, without waiting.
    pub fn drain(&self) -> impl Iterator<Item = Change> + '_ {
        self.changes.try_iter()
    }

    /// Wait up to `within` for the next change.
    #[must_use]
    pub fn next_change(&self, within: Duration) -> Option<Change> {
        self.changes.recv_timeout(within).ok()
    }
}

/// Whether the target application has anything of its own in the content
/// directory.
///
/// Asked of the directory rather than tracked from events, because events can
/// be missed and the answer has to be right at the moment somebody is deciding
/// whether to close a session (concept 6.2).
///
/// # Errors
///
/// Where the directory cannot be read.
pub fn siblings_present(dir: &Path, content_name: &str) -> std::io::Result<bool> {
    for entry in std::fs::read_dir(dir)? {
        if entry?.file_name() != *content_name {
            return Ok(true);
        }
    }
    Ok(false)
}

/// **The stop is waited for, and this is the fix for a measured hang.**
///
/// `notify`'s watcher drop is fire-and-forget: it posts `Action::Stop`, wakes
/// its server thread and returns, leaving that thread inside `stop_watch`.
/// Whatever removes the watched directory next therefore races a watch that is
/// still stopping, and on Windows that pair deadlocks — `stop_watch` waits
/// `INFINITE` for a semaphore its own completion routine does not post when it
/// re-arms `ReadDirectoryChangesW`, and the re-armed read does not return while
/// a `remove_dir_all` is walking the same directory.
///
/// Measured 2026-09-07 before this existed: a diagnostic that drops a watch and
/// then removes the directory wedged 9 runs in 20, one instance parked with
/// zero CPU for over ten minutes; `Opened::close`, which removes first and
/// drops after, wedged past a 300-second timeout with the same two stacks.
/// `docs/windows-save-test-hang.md` has both, and the ordering alone was never
/// the cure: the 9-in-20 case already dropped the watch first.
///
/// So the watcher goes, and then this waits for the backend to say the watch
/// has actually stopped, which is the guarantee the caller needs before
/// removing anything. [`STOP_WAIT`] bounds the wait so that a stop which never
/// finishes is a delay rather than a second hang.
impl Drop for Watch {
    fn drop(&mut self) {
        drop(self.watcher.take());

        #[cfg(windows)]
        {
            // The same channel carries `WatcherAwakened`, so this reads until
            // the completion it wants, the sender goes, or the clock runs out.
            let deadline = std::time::Instant::now() + STOP_WAIT;
            loop {
                let left = deadline.saturating_duration_since(std::time::Instant::now());
                if left.is_zero() {
                    break;
                }
                match self.stopped.recv_timeout(left) {
                    Ok(notify::windows::MetaEvent::SingleWatchComplete) | Err(_) => break,
                    Ok(_) => {}
                }
            }
        }
    }
}

#[cfg(test)]
mod tests {
    use super::{classify, siblings_present, Change, EventKind, Watch};
    use std::path::{Path, PathBuf};
    use std::time::Duration;

    fn at(names: &[&str]) -> Vec<PathBuf> {
        names
            .iter()
            .map(|n| Path::new("/s/content").join(n))
            .collect()
    }

    fn refs(paths: &[PathBuf]) -> Vec<&Path> {
        paths.iter().map(AsRef::as_ref).collect()
    }

    #[test]
    fn writing_the_content_file_is_the_write_back_trigger() {
        let p = at(&["report.pdf"]);
        assert_eq!(
            classify("report.pdf", &refs(&p), EventKind::Touched),
            [Change::Content]
        );
    }

    #[test]
    fn anything_else_appearing_is_the_application_working() {
        // No table of lock file conventions. The directory held one file, this
        // tool put it there, so whatever this is came from the editor.
        for name in [
            "~$report.docx",
            ".~lock.report.pdf#",
            "report.pdf.tmp",
            "4919",
        ] {
            let p = at(&[name]);
            assert_eq!(
                classify("report.pdf", &refs(&p), EventKind::Touched),
                [Change::SiblingAppeared],
                "{name}"
            );
        }
    }

    #[test]
    fn a_sibling_going_away_is_the_application_finishing() {
        let p = at(&["~$report.docx"]);
        assert_eq!(
            classify("report.pdf", &refs(&p), EventKind::Gone),
            [Change::SiblingWentAway]
        );
    }

    #[test]
    fn the_content_file_going_away_is_still_the_content_file() {
        // A rename over it arrives as the content file being replaced, and an
        // application that deletes and rewrites is doing a save in two steps.
        // Either way the container should be asked to catch up.
        let p = at(&["report.pdf"]);
        assert_eq!(
            classify("report.pdf", &refs(&p), EventKind::Gone),
            [Change::Content]
        );
    }

    #[test]
    fn a_rename_naming_both_paths_reports_both() {
        // The atomic save: a temporary sibling renamed over the target. The
        // sibling says the application is working and the content file is the
        // save, and dropping either would lose one of the two things the watch
        // is for.
        let p = at(&["report.pdf.tmp", "report.pdf"]);
        assert_eq!(
            classify("report.pdf", &refs(&p), EventKind::Touched),
            [Change::SiblingAppeared, Change::Content]
        );
    }

    #[test]
    fn a_content_file_named_like_a_lock_file_is_still_the_content_file() {
        // SPEC 2.3 permits any plain filename. Matching by name and not by
        // shape is what keeps this true.
        let p = at(&["~$report.docx"]);
        assert_eq!(
            classify("~$report.docx", &refs(&p), EventKind::Touched),
            [Change::Content]
        );
    }

    #[test]
    fn siblings_are_asked_of_the_directory_rather_than_remembered() {
        let tmp = tempfile::tempdir().unwrap();
        std::fs::write(tmp.path().join("report.pdf"), b"x").unwrap();
        assert!(!siblings_present(tmp.path(), "report.pdf").unwrap());

        std::fs::write(tmp.path().join("~$report.pdf"), b"").unwrap();
        assert!(siblings_present(tmp.path(), "report.pdf").unwrap());

        std::fs::remove_file(tmp.path().join("~$report.pdf")).unwrap();
        assert!(!siblings_present(tmp.path(), "report.pdf").unwrap());
    }

    #[test]
    fn reading_the_content_file_is_not_a_change_to_it() {
        // Write-back opens the content file to read it, which inotify reports
        // as an access on the watched file. Treating that as a save makes the
        // write-back its own trigger: measured on 2026-08-30, one edit produced
        // three repacks and would have produced more had the session stayed
        // open.
        //
        // **The read has to be told apart from the setup, and only after the
        // watch is quiet.** FSEvents is path-based and replays: registering a
        // watch delivers the events that just happened to the directory,
        // including the tempdir appearing and the `write` below, at the
        // platform's own latency rather than before this returns. Measured on
        // an Apple silicon runner 2026-09-07, the third run in a loop: the
        // setup arrived as `[SiblingAppeared, Content, Content]` inside the
        // window and read as the read. inotify and ReadDirectoryChangesW
        // deliver only what follows registration and never showed it. So the
        // watch is drained until it goes quiet, which absorbs the replay on
        // every platform, and only what the read then produces is measured —
        // which is nothing: probed five times on that platform, a settled read
        // produced no `Content`, because a read is `Access` and `EventKind::of`
        // discards it. The earlier version had no settle and asserted against
        // the whole window, so it was measuring the notifier.
        let tmp = tempfile::tempdir().unwrap();
        let content_path = tmp.path().join("report.pdf");
        std::fs::write(&content_path, b"first").unwrap();

        let watch = Watch::on(tmp.path(), "report.pdf").unwrap();

        // Quiet is nothing for 400ms, capped so a notifier that never falls
        // silent cannot hang the suite. The replay is what is being waited out.
        let cap = std::time::Instant::now() + Duration::from_secs(5);
        while watch.next_change(Duration::from_millis(400)).is_some()
            && std::time::Instant::now() < cap
        {}

        let _ = std::fs::read(&content_path).unwrap();

        let deadline = std::time::Instant::now() + Duration::from_secs(2);
        let mut seen = Vec::new();
        while std::time::Instant::now() < deadline {
            if let Some(c) = watch.next_change(Duration::from_millis(100)) {
                seen.push(c);
            }
        }
        assert!(
            !seen.contains(&Change::Content),
            "reading the content file was reported as a change: {seen:?}"
        );
    }

    #[test]
    fn a_real_atomic_save_reaches_the_watch() {
        // The one test that goes through the platform. It saves the way a
        // serious editor does — write a temporary sibling, rename over the
        // target — which is the case a watch registered on the file would miss
        // entirely.
        let tmp = tempfile::tempdir().unwrap();
        let content_path = tmp.path().join("report.pdf");
        std::fs::write(&content_path, b"first").unwrap();

        let watch = Watch::on(tmp.path(), "report.pdf").unwrap();

        let scratch = tmp.path().join("report.pdf.tmp");
        std::fs::write(&scratch, b"second").unwrap();
        std::fs::rename(&scratch, &content_path).unwrap();

        // Generously, because this is at the platform's pace and not ours.
        let mut seen = Vec::new();
        let deadline = std::time::Instant::now() + Duration::from_secs(10);
        while std::time::Instant::now() < deadline && !seen.contains(&Change::Content) {
            if let Some(c) = watch.next_change(Duration::from_millis(250)) {
                seen.push(c);
            }
        }
        assert!(
            seen.contains(&Change::Content),
            "the save never arrived: {seen:?}"
        );
        assert_eq!(std::fs::read(&content_path).unwrap(), b"second");
    }
}