treeship_core/session/

event_log.rs

//! Append-only, file-backed event log for session events.
//!
//! Events are stored as newline-delimited JSON (JSONL) in
//! `.treeship/sessions/<session_id>/events.jsonl`.
//!
//! Concurrency model: `append()` is safe to call from multiple processes
//! concurrently. Each call attempts to acquire an exclusive advisory lock
//! (via `fs2::FileExt::try_lock_exclusive` -- backed by `flock(2)` on Unix
//! and `LockFileEx` on Windows) on a sidecar `events.jsonl.lock` file in a
//! ~500ms bounded retry loop. Under the lock, a counter sidecar
//! `events.jsonl.count` is the authoritative source for the next
//! `sequence_no`. The per-process AtomicU64 is retained as a hot-path
//! optimization for non-contended use, but its value is overwritten by the
//! on-disk counter after every locked append.
//!
//! Counter sidecar format (16 bytes):
//!   - bytes 0..8:  count (u64 LE) -- number of events written to events.jsonl
//!   - bytes 8..16: byte_size (u64 LE) -- size of events.jsonl when count was recorded
//!
//! The byte_size field is the crash detector. If a peer wrote events.jsonl
//! but crashed before fsyncing the counter (or vice versa), the size on disk
//! and the size in the counter disagree. On any mismatch we fall back to an
//! O(N) line count and rewrite the counter -- one paid scan, then back to
//! O(1) on every subsequent append.
//!
//! This bounds steady-state append cost at constant: read 16 bytes, write
//! one JSONL line, write 16 bytes. The previous implementation re-streamed
//! the entire events.jsonl on every append, which made hooks O(N) in
//! session length and dominated PostToolUse latency on long sessions.
//!
//! Fail-closed semantics: the writer ALWAYS acquires the exclusive flock
//! before reading the counter and writing the event. We use fs2's blocking
//! `lock_exclusive()` (flock(2) without LOCK_NB on Unix, LockFileEx without
//! LOCKFILE_FAIL_IMMEDIATELY on Windows) so contended writers queue rather
//! than race. An earlier "best-effort, fall through on contention" path
//! existed here -- it could produce duplicate `sequence_no` under hook
//! contention (P0, audit lane F), so it was removed. Hook callers already
//! sit under Claude Code's 60s hook timeout, which bounds wall-clock
//! exposure to a wedged peer. Trading a bounded wait for guaranteed
//! injective sequence numbers is the right call when receipts are the
//! trust artifact.
//!
//! Lock file permissions are 0o600 (owner-only) on Unix, applied at file
//! creation via `OpenOptionsExt::mode` and re-tightened on every open if
//! a previous run left the file with looser perms.

use std::io::{BufRead, Write};
use std::path::{Path, PathBuf};
use std::sync::atomic::{AtomicU64, Ordering};

#[cfg(not(target_family = "wasm"))]
use fs2::FileExt;

use crate::session::event::SessionEvent;

/// Error from event log operations.
#[derive(Debug)]
pub enum EventLogError {
    Io(std::io::Error),
    Json(serde_json::Error),
}

impl std::fmt::Display for EventLogError {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        match self {
            Self::Io(e) => write!(f, "event log io: {e}"),
            Self::Json(e) => write!(f, "event log json: {e}"),
        }
    }
}

impl std::error::Error for EventLogError {}
impl From<std::io::Error> for EventLogError {
    fn from(e: std::io::Error) -> Self {
        Self::Io(e)
    }
}
impl From<serde_json::Error> for EventLogError {
    fn from(e: serde_json::Error) -> Self {
        Self::Json(e)
    }
}

/// An append-only event log backed by a JSONL file.
pub struct EventLog {
    path: PathBuf,
    sequence: AtomicU64,
}

impl EventLog {
    /// Open or create an event log for the given session directory.
    ///
    /// The session directory is typically `.treeship/sessions/<session_id>/`.
    /// If the directory does not exist, it will be created.
    ///
    /// Initialization reads the counter sidecar in O(1) when present and
    /// consistent with events.jsonl's byte size; falls back to an O(N) line
    /// count (and rewrites the sidecar) when the sidecar is missing,
    /// short-read, or stale from a crashed previous appender.
    pub fn open(session_dir: &Path) -> Result<Self, EventLogError> {
        std::fs::create_dir_all(session_dir)?;
        let path = session_dir.join("events.jsonl");
        // Read-only. `open` used to call `read_counter_or_recount`, which
        // REWRITES the counter sidecar when it finds it stale or missing --
        // an unlocked read-modify-write of the same shared state
        // `append_locked` takes an exclusive flock to protect.
        //
        // In the model that matters, every hook invocation constructs a fresh
        // handle, so `open` sits inside the contention window: one process
        // could clobber the counter with a value computed before another
        // process's append landed. The size check in
        // `read_counter_consistent` makes that self-healing rather than
        // fatal, but a constructor writing shared state behind the lock's
        // back is a discipline violation regardless of whether a given
        // interleaving is survivable. Issue #275.
        //
        // Locking `open` instead would serialize every handle construction
        // on an exclusive flock just to seed a hint. Not writing is both
        // cheaper and more obviously correct: on native, this value is only
        // a hint. `append_locked` re-reads the counter INSIDE the lock and
        // overwrites `sequence` (see the store at the end of that function),
        // and counter repair still happens there, where it is serialized.
        let count = read_counter_hint(&path);
        Ok(Self {
            path,
            sequence: AtomicU64::new(count),
        })
    }

    /// Append a single event to the log.
    ///
    /// The event's `sequence_no` is set automatically. Under contention from
    /// multiple writer processes, the sequence number is re-derived from the
    /// on-disk line count under an exclusive flock so two parallel writers
    /// never collide.
    pub fn append(&self, event: &mut SessionEvent) -> Result<(), EventLogError> {
        self.append_locked(event)
    }

    /// Cross-process safe append: acquires an exclusive advisory lock on a
    /// sidecar `.lock` file, re-counts events.jsonl lines, assigns sequence_no,
    /// writes the new event, then releases the lock on drop.
    ///
    /// Locking is BLOCKING (`fs2::FileExt::lock_exclusive`, i.e. flock(2)
    /// without LOCK_NB). A previous implementation polled `try_lock_exclusive`
    /// for ~500ms and then fell through to an UNLOCKED write -- which under
    /// hook contention (multiple PostToolUse invocations racing) could
    /// assign duplicate `sequence_no` values to different events. That
    /// broke the injective-sequence invariant receipts depend on, even
    /// though local merkle verification still passed. Audit lane F (P0).
    ///
    /// The trade-off: a wedged peer holding the lock will now stall this
    /// caller until the peer releases (e.g. by crashing -- flock is
    /// released by the kernel when the holder's FD closes). Hook
    /// invocations are bounded by Claude Code's hook timeout (60s by
    /// default), so the worst-case wedge surfaces as a hook failure
    /// rather than a duplicate sequence_no. That mirrors how
    /// `journal/mod.rs` treats the journal append lock as a hard
    /// correctness barrier rather than a soft hint.
    ///
    /// The locked region covers BOTH the counter read AND the file write,
    /// so two concurrent writers cannot read the same count and append
    /// twice with that count.
    ///
    /// Lock file is created mode 0o600 (owner-only) so the sidecar can
    /// never be opened by other users on a shared machine.
    ///
    /// Skipped on WASM (no fs, no concurrency).
    #[cfg(not(target_family = "wasm"))]
    fn append_locked(&self, event: &mut SessionEvent) -> Result<(), EventLogError> {
        // Sidecar lock file: contention here doesn't block readers of events.jsonl.
        let lock_path = self.path.with_extension("jsonl.lock");

        // Open or create the lock file. On Unix we set 0o600 explicitly so
        // the sidecar isn't group/world readable; the umask-derived default
        // would otherwise be permissive on some setups.
        let lock_file = open_lock_file(&lock_path)?;

        // Blocking flock. Returns when the lock is held exclusively, or
        // propagates a real I/O error (filesystem that doesn't support
        // flock, FD revoked, etc.). EINTR retry isn't needed -- fs2 wraps
        // the syscall and retries internally on POSIX.
        FileExt::lock_exclusive(&lock_file)?;

        // From here until `lock_file` is dropped (end of function), we
        // hold the exclusive flock. The counter read + event write + counter
        // update MUST stay inside this block; any early return must still
        // drop `lock_file`, which Rust guarantees by RAII.
        let result = (|| -> Result<(), EventLogError> {
            // Read sequence_no from the counter sidecar in O(1) when
            // consistent with events.jsonl size. Stale or missing counters
            // force a one-time O(N) rescan that also rewrites the counter,
            // so subsequent appends return to O(1). Only the on-disk state
            // (counter + size check) is authoritative when multiple
            // processes are appending; the per-process AtomicU64 is a
            // stale hint.
            let count = read_counter_or_recount(&self.path)?;
            event.sequence_no = count;

            let mut line = serde_json::to_vec(event)?;
            line.push(b'\n');

            let mut file = std::fs::OpenOptions::new()
                .create(true)
                .append(true)
                .open(&self.path)?;
            file.write_all(&line)?;
            file.flush()?;

            // Update the counter sidecar with the new count and the new
            // events.jsonl size, so the next append can short-circuit the
            // line scan. Failure to update the counter is non-fatal: the
            // next reader will detect the size mismatch and recount.
            let new_size = file.metadata().map(|m| m.len()).unwrap_or(0);
            let _ = write_counter(&self.path, count + 1, new_size);

            // Keep the in-process AtomicU64 in sync so non-contended callers
            // see the right value via event_count() without re-reading.
            self.sequence.store(count + 1, Ordering::SeqCst);
            Ok(())
        })();

        // Explicit unlock matches the journal precedent (journal/mod.rs
        // also calls `unlock` before dropping). Drop alone would release
        // the flock via close(2), but being explicit makes the lock
        // window obvious to readers of this function.
        let _ = FileExt::unlock(&lock_file);
        result
    }

    /// WASM build: no filesystem locks available, no concurrent writers.
    /// Falls back to the simple AtomicU64 path.
    #[cfg(target_family = "wasm")]
    fn append_locked(&self, event: &mut SessionEvent) -> Result<(), EventLogError> {
        event.sequence_no = self.sequence.fetch_add(1, Ordering::SeqCst);

        let mut line = serde_json::to_vec(event)?;
        line.push(b'\n');

        let mut file = std::fs::OpenOptions::new()
            .create(true)
            .append(true)
            .open(&self.path)?;
        file.write_all(&line)?;
        file.flush()?;

        Ok(())
    }

    /// Read all events from the log.
    ///
    /// Per-line tolerant: a single bad line -- malformed JSON (unknown
    /// event type, missing field, truncated write) OR a non-UTF-8 byte
    /// (partial write, corruption) -- is logged to stderr, counted as a
    /// skip, and stepped over, not propagated as an error. The caller --
    /// session close, in particular -- composes a receipt from whatever
    /// events parse, instead of dropping every event when any one is bad.
    /// (The read only returns Err for a whole-file failure such as the
    /// events file being unopenable; a per-line decode error is a skip.)
    ///
    /// Why this matters: events.jsonl is append-only and written by
    /// hooks, daemons, SDKs, and bridges from multiple processes. A
    /// single bad event from one buggy emitter would otherwise nuke
    /// the entire receipt's side_effects / agent_graph / timeline.
    /// Real-world repro: a hook that emitted events with an unknown
    /// `type` field caused side_effects.files_written to come back
    /// empty even though the rest of the events in the log were valid
    /// agent.wrote_file events the aggregator would have happily
    /// processed.
    pub fn read_all(&self) -> Result<Vec<SessionEvent>, EventLogError> {
        // Drop the skipped count for callers that don't carry it through.
        // Receipt composition uses read_all_with_stats to record the
        // count in-band on the sealed receipt -- see Codex finding #8.
        self.read_all_with_stats().map(|(events, _skipped)| events)
    }

    /// Same as `read_all` but returns the count of malformed lines that
    /// were skipped during parsing alongside the valid events.
    ///
    /// Codex adversarial review finding #8: skipping malformed events
    /// on stderr only is silent data loss from the verifier's
    /// perspective. The receipt gets sealed under a merkle root that
    /// represents only the events that successfully parsed -- a
    /// downstream consumer cannot tell whether the receipt is complete
    /// or whether N events were silently dropped.
    ///
    /// session::close calls this and stores the count on
    /// `receipt.proofs.event_log_skipped`. `treeship package verify`
    /// surfaces it as a WARN when nonzero so the receipt's
    /// completeness signal is visible without breaking byte-identical
    /// re-verification of pre-existing receipts.
    pub fn read_all_with_stats(&self) -> Result<(Vec<SessionEvent>, usize), EventLogError> {
        if !self.path.exists() {
            return Ok((Vec::new(), 0));
        }
        let file = std::fs::File::open(&self.path)?;
        let reader = std::io::BufReader::new(file);
        let mut events = Vec::new();
        let mut skipped = 0usize;
        for (idx, line) in reader.lines().enumerate() {
            // A per-LINE read error (a non-UTF-8 byte from a partial write or
            // corruption) must be counted as a skip, not propagated — the `?`
            // here previously aborted the WHOLE read, and `session::close`
            // maps that Err to (empty, 0), sealing an empty receipt with a
            // zero skip-count exactly when the log is most damaged. That
            // silently bypasses the completeness signal this function exists
            // to carry. Treat the bad line like a malformed JSON line: skip
            // it, count it, keep the rest.
            let line = match line {
                Ok(l) => l,
                Err(e) => {
                    skipped += 1;
                    eprintln!(
                        "[treeship] event_log: skipping unreadable line {} in {}: {}",
                        idx + 1,
                        self.path.display(),
                        e,
                    );
                    continue;
                }
            };
            if line.trim().is_empty() {
                continue;
            }
            match serde_json::from_str::<SessionEvent>(&line) {
                Ok(event) => events.push(event),
                Err(e) => {
                    skipped += 1;
                    eprintln!(
                        "[treeship] event_log: skipping malformed line {} in {}: {}",
                        idx + 1,
                        self.path.display(),
                        e,
                    );
                }
            }
        }
        if skipped > 0 {
            eprintln!(
                "[treeship] event_log: {} malformed line(s) skipped while reading {} (kept {} valid event(s))",
                skipped,
                self.path.display(),
                events.len(),
            );
        }
        Ok((events, skipped))
    }

    /// Return the current event count.
    pub fn event_count(&self) -> u64 {
        self.sequence.load(Ordering::SeqCst)
    }

    /// Return the path to the JSONL file.
    pub fn path(&self) -> &Path {
        &self.path
    }
}

/// Open the sidecar lock file with owner-only permissions (0o600 on Unix).
///
/// On Unix the mode is set atomically via `OpenOptionsExt::mode` for newly
/// created files. For files that already exist (e.g. left over from a
/// prior crash or an upgrade from a pre-0.9.3 CLI that didn't tighten
/// perms), we additionally re-chmod to 0o600 after open IF the file is
/// owned by the current user. This is best-effort: if the chmod fails
/// (file owned by another user, read-only filesystem, etc.) we proceed
/// silently rather than refuse to open the lock -- the lock semantics
/// don't depend on the perms being tight, only the privacy of the
/// sidecar's existence does.
///
/// On Windows the mode concept doesn't apply; ACLs default to inheriting
/// the parent dir's permissions, which for `.treeship/sessions/<id>/`
/// should already be scoped to the owning user.
#[cfg(all(not(target_family = "wasm"), unix))]
fn open_lock_file(path: &Path) -> Result<std::fs::File, std::io::Error> {
    use std::os::unix::fs::{MetadataExt, OpenOptionsExt, PermissionsExt};
    use std::os::unix::io::AsRawFd;

    let file = std::fs::OpenOptions::new()
        .create(true)
        .read(true)
        .write(true)
        .mode(0o600)
        .open(path)?;

    // Re-tighten if a pre-existing file has loose perms. Use `fchmod` on the
    // open file descriptor rather than `set_permissions(path, ...)` to
    // eliminate the TOCTOU window -- between metadata() and a path-based
    // chmod, an attacker could swap the file. `fchmod` operates on the
    // already-opened inode, so the target is pinned.
    //
    // Only act when the file is owned by us (uid match via geteuid). If
    // fchmod fails (NFS mount with restricted metadata writes, or some
    // filesystems without full POSIX perm support), emit a one-line
    // stderr warning so an operator has visibility. The lock still works;
    // only the privacy of the sidecar's existence is affected.
    if let Ok(meta) = file.metadata() {
        let mode = meta.permissions().mode() & 0o777;
        let owned_by_us = meta.uid() == nix_uid();
        if owned_by_us && mode != 0o600 {
            let fd = file.as_raw_fd();
            // SAFETY: fd is valid (we just opened it), 0o600 is a
            // well-formed mode. fchmod is async-signal-safe per POSIX.
            let rc = unsafe { libc_fchmod(fd, 0o600) };
            if rc != 0 {
                let err = std::io::Error::last_os_error();
                eprintln!(
                    "[treeship] warning: could not tighten lock file perms on {} \
                     to 0o600 (current: 0o{:o}). Error: {}. Lock still functions; \
                     only the privacy of the sidecar is affected. Common cause: \
                     NFS mount or filesystem without full POSIX perm support.",
                    path.display(),
                    mode,
                    err
                );
            }
        }
    }

    Ok(file)
}

/// Thin FFI wrapper around libc::fchmod. Declared here so event_log.rs
/// doesn't need a direct libc crate dep -- the symbol is available in
/// every Unix libc binary.
#[cfg(all(not(target_family = "wasm"), unix))]
fn libc_fchmod(fd: i32, mode: u32) -> i32 {
    // SAFETY: posix-standard FFI signature; `fd` validity and `mode`
    // bounds are enforced by the caller.
    unsafe extern "C" {
        fn fchmod(fd: i32, mode: u32) -> i32;
    }
    unsafe { fchmod(fd, mode) }
}

/// Lightweight wrapper around `geteuid` so we can compare to file ownership
/// without pulling in the `nix` crate. Uses `libc` directly (already a
/// transitive dep via several upstream crates).
#[cfg(all(not(target_family = "wasm"), unix))]
fn nix_uid() -> u32 {
    // SAFETY: geteuid is async-signal-safe and never fails per POSIX.
    unsafe extern "C" {
        fn geteuid() -> u32;
    }
    unsafe { geteuid() }
}

#[cfg(all(not(target_family = "wasm"), not(unix)))]
fn open_lock_file(path: &Path) -> Result<std::fs::File, std::io::Error> {
    std::fs::OpenOptions::new()
        .create(true)
        .read(true)
        .write(true)
        .open(path)
}

/// Path of the counter sidecar for a given events.jsonl path.
fn counter_path(events_path: &Path) -> PathBuf {
    events_path.with_extension("jsonl.count")
}

/// Read the counter sidecar if it exists and is consistent with events.jsonl.
///
/// Returns `Some(count)` when the sidecar's recorded byte_size matches the
/// current events.jsonl size, and `None` otherwise (missing sidecar, short
/// read, parse failure, or size mismatch from a crashed previous appender).
#[cfg(not(target_family = "wasm"))]
fn read_counter_consistent(events_path: &Path) -> Option<u64> {
    let counter = counter_path(events_path);
    let bytes = std::fs::read(&counter).ok()?;
    if bytes.len() != 16 {
        return None;
    }
    let count = u64::from_le_bytes(bytes[0..8].try_into().ok()?);
    let recorded_size = u64::from_le_bytes(bytes[8..16].try_into().ok()?);

    // events.jsonl may not exist yet -- counter records (0, 0) for that case.
    let actual_size = match std::fs::metadata(events_path) {
        Ok(m) => m.len(),
        Err(e) if e.kind() == std::io::ErrorKind::NotFound => 0,
        Err(_) => return None,
    };
    if actual_size != recorded_size {
        return None;
    }
    Some(count)
}

/// Read the counter via the sidecar (O(1)) or fall back to an O(N) line
/// scan, rewriting the sidecar on the way out. This is the recovery path
/// after a crash that left the counter and events.jsonl out of sync.
#[cfg(not(target_family = "wasm"))]
/// Best-effort event count for seeding the in-process hint, with NO writes.
///
/// Used by `open`. Never repairs the counter sidecar: repair belongs inside
/// the append lock, where it cannot race a concurrent writer. A wrong hint is
/// harmless on native -- `append_locked` re-reads authoritatively under the
/// lock before assigning any `sequence_no`.
#[cfg(not(target_family = "wasm"))]
fn read_counter_hint(events_path: &Path) -> u64 {
    if let Some(count) = read_counter_consistent(events_path) {
        return count;
    }
    let Ok(f) = std::fs::File::open(events_path) else {
        return 0;
    };
    std::io::BufReader::new(f)
        .lines()
        .filter(|l| l.is_ok())
        .count() as u64
}

/// WASM: no fs, no concurrent writers, so the in-memory AtomicU64 is the
/// whole story. There is no wasm `read_counter_or_recount` counterpart --
/// its only caller is `append_locked`, which is itself native-only.
#[cfg(target_family = "wasm")]
fn read_counter_hint(_events_path: &Path) -> u64 {
    0
}

#[cfg(not(target_family = "wasm"))]
fn read_counter_or_recount(events_path: &Path) -> Result<u64, EventLogError> {
    if let Some(count) = read_counter_consistent(events_path) {
        return Ok(count);
    }
    let count = if events_path.exists() {
        let f = std::fs::File::open(events_path)?;
        let r = std::io::BufReader::new(f);
        r.lines().filter(|l| l.is_ok()).count() as u64
    } else {
        0
    };
    let size = std::fs::metadata(events_path).map(|m| m.len()).unwrap_or(0);
    let _ = write_counter(events_path, count, size);
    Ok(count)
}

/// Atomically replace the counter sidecar with the new (count, byte_size).
///
/// Writes to a temp file in the same directory and renames into place so a
/// reader either sees the old 16 bytes or the new 16 bytes, never a partial
/// write. The 0o600 perm matches the lock file -- the counter doesn't leak
/// secrets but its existence is a session signal worth scoping to the owner.
#[cfg(not(target_family = "wasm"))]
fn write_counter(events_path: &Path, count: u64, byte_size: u64) -> Result<(), std::io::Error> {
    use std::io::Write as _;
    let counter = counter_path(events_path);
    let dir = counter.parent().ok_or_else(|| {
        std::io::Error::new(
            std::io::ErrorKind::InvalidInput,
            "counter path has no parent",
        )
    })?;
    std::fs::create_dir_all(dir)?;

    let mut buf = [0u8; 16];
    buf[0..8].copy_from_slice(&count.to_le_bytes());
    buf[8..16].copy_from_slice(&byte_size.to_le_bytes());

    let tmp = counter.with_extension("count.tmp");
    {
        let mut f = open_counter_tmp(&tmp)?;
        f.write_all(&buf)?;
        f.sync_all()?;
    }
    std::fs::rename(&tmp, &counter)?;
    Ok(())
}

#[cfg(all(not(target_family = "wasm"), unix))]
fn open_counter_tmp(path: &Path) -> Result<std::fs::File, std::io::Error> {
    use std::os::unix::fs::OpenOptionsExt;
    std::fs::OpenOptions::new()
        .create(true)
        .write(true)
        .truncate(true)
        .mode(0o600)
        .open(path)
}

#[cfg(all(not(target_family = "wasm"), not(unix)))]
fn open_counter_tmp(path: &Path) -> Result<std::fs::File, std::io::Error> {
    std::fs::OpenOptions::new()
        .create(true)
        .write(true)
        .truncate(true)
        .open(path)
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::session::event::*;

    pub(super) fn make_event(session_id: &str, event_type: EventType) -> SessionEvent {
        SessionEvent {
            session_id: session_id.into(),
            event_id: generate_event_id(),
            timestamp: "2026-04-05T08:00:00Z".into(),
            sequence_no: 0,
            trace_id: generate_trace_id(),
            span_id: generate_span_id(),
            parent_span_id: None,
            agent_id: "agent://test".into(),
            agent_instance_id: "ai_test_1".into(),
            agent_name: "test-agent".into(),
            agent_role: None,
            host_id: "host_test".into(),
            tool_runtime_id: None,
            event_type,
            artifact_ref: None,
            meta: None,
        }
    }

    #[test]
    fn append_and_read_back() {
        let dir =
            std::env::temp_dir().join(format!("treeship-evtlog-test-{}", rand::random::<u32>()));
        let log = EventLog::open(&dir).unwrap();

        let mut e1 = make_event("ssn_001", EventType::SessionStarted);
        let mut e2 = make_event(
            "ssn_001",
            EventType::AgentStarted {
                parent_agent_instance_id: None,
            },
        );

        log.append(&mut e1).unwrap();
        log.append(&mut e2).unwrap();

        assert_eq!(log.event_count(), 2);
        assert_eq!(e1.sequence_no, 0);
        assert_eq!(e2.sequence_no, 1);

        let events = log.read_all().unwrap();
        assert_eq!(events.len(), 2);
        assert_eq!(events[0].sequence_no, 0);
        assert_eq!(events[1].sequence_no, 1);

        let _ = std::fs::remove_dir_all(&dir);
    }

    #[test]
    fn read_all_skips_malformed_lines() {
        // Regression: a single malformed line in events.jsonl used to
        // make read_all() return Err, and the caller's
        // .unwrap_or_default() would drop EVERY event in the log. Real
        // bug: hooks emitting events with an unknown `type` field made
        // side_effects.files_written come back empty even though every
        // other event in the log was a perfectly valid agent.wrote_file
        // event. Now we skip-and-log the bad line and keep the rest.
        let dir = std::env::temp_dir().join(format!(
            "treeship-evtlog-malformed-{}",
            rand::random::<u32>()
        ));
        let log = EventLog::open(&dir).unwrap();

        let mut good1 = make_event(
            "ssn_001",
            EventType::AgentWroteFile {
                file_path: "src/before.rs".into(),
                digest: None,
                operation: None,
                additions: None,
                deletions: None,
            },
        );
        let mut good2 = make_event(
            "ssn_001",
            EventType::AgentWroteFile {
                file_path: "src/after.rs".into(),
                digest: None,
                operation: None,
                additions: None,
                deletions: None,
            },
        );
        log.append(&mut good1).unwrap();
        log.append(&mut good2).unwrap();

        // Manually inject a malformed line between the two good ones by
        // truncating the file and rewriting. The malformed line has an
        // unknown event type ("custom.weird") which the closed EventType
        // enum can't deserialize.
        let path = log.path().to_path_buf();
        let original = std::fs::read_to_string(&path).unwrap();
        let mut lines: Vec<&str> = original.lines().collect();
        lines.insert(1, r#"{"session_id":"ssn_001","event_id":"evt_bad","timestamp":"2026-04-26T00:00:00Z","sequence_no":1,"trace_id":"x","span_id":"y","agent_id":"a","agent_instance_id":"i","agent_name":"n","host_id":"h","type":"custom.weird","payload":42}"#);
        std::fs::write(&path, lines.join("\n") + "\n").unwrap();

        let events = log.read_all().unwrap();
        assert_eq!(
            events.len(),
            2,
            "expected the two valid events to come through; got {}",
            events.len()
        );
        // Confirm the valid events are the file-write events and not
        // some default fallback.
        let written_paths: Vec<&str> = events
            .iter()
            .filter_map(|e| match &e.event_type {
                EventType::AgentWroteFile { file_path, .. } => Some(file_path.as_str()),
                _ => None,
            })
            .collect();
        assert_eq!(written_paths, vec!["src/before.rs", "src/after.rs"]);

        // Codex finding #8: the count must be exposed in-band so a
        // sealed receipt can carry the incompleteness signal. Verify
        // read_all_with_stats reports it.
        let (events2, skipped) = log.read_all_with_stats().unwrap();
        assert_eq!(events2.len(), 2);
        assert_eq!(
            skipped, 1,
            "exactly one malformed line was injected; expected skipped == 1"
        );

        let _ = std::fs::remove_dir_all(&dir);
    }

    #[test]
    fn read_all_with_stats_reports_zero_when_clean() {
        // No malformed lines -> skipped == 0 -> the receipt's
        // event_log_skipped field stays default (0) and gets omitted
        // from canonical JSON. This preserves byte-identical receipts
        // for the common case where the event log is clean.
        let dir =
            std::env::temp_dir().join(format!("treeship-evtlog-clean-{}", rand::random::<u32>()));
        let log = EventLog::open(&dir).unwrap();

        let mut e = make_event(
            "ssn_001",
            EventType::AgentWroteFile {
                file_path: "x.rs".into(),
                digest: None,
                operation: None,
                additions: None,
                deletions: None,
            },
        );
        log.append(&mut e).unwrap();

        let (events, skipped) = log.read_all_with_stats().unwrap();
        assert_eq!(events.len(), 1);
        assert_eq!(skipped, 0);

        let _ = std::fs::remove_dir_all(&dir);
    }

    #[test]
    fn non_utf8_byte_is_skipped_not_aborted() {
        // A single non-UTF-8 byte in the log (partial write / corruption)
        // must skip THAT line and keep the rest, with a nonzero skip count —
        // NOT abort the whole read and seal an empty receipt with skipped=0.
        use std::io::Write;
        let dir =
            std::env::temp_dir().join(format!("treeship-evtlog-badbyte-{}", rand::random::<u32>()));
        let log = EventLog::open(&dir).unwrap();
        let mut e = make_event(
            "ssn_001",
            EventType::AgentWroteFile {
                file_path: "ok.rs".into(),
                digest: None,
                operation: None,
                additions: None,
                deletions: None,
            },
        );
        log.append(&mut e).unwrap();

        // Append a raw non-UTF-8 line directly to the events file.
        let events_path = dir.join("events.jsonl");
        let mut f = std::fs::OpenOptions::new()
            .append(true)
            .open(&events_path)
            .unwrap();
        f.write_all(&[0xff, 0xfe, b'\n']).unwrap(); // invalid UTF-8 line
        drop(f);

        let (events, skipped) = log.read_all_with_stats().unwrap();
        assert_eq!(events.len(), 1, "the one good event must survive");
        assert_eq!(
            skipped, 1,
            "the bad line must be counted as skipped, not silently dropped"
        );

        let _ = std::fs::remove_dir_all(&dir);
    }

    #[test]
    fn reopen_preserves_sequence() {
        let dir =
            std::env::temp_dir().join(format!("treeship-evtlog-reopen-{}", rand::random::<u32>()));

        {
            let log = EventLog::open(&dir).unwrap();
            let mut e = make_event("ssn_001", EventType::SessionStarted);
            log.append(&mut e).unwrap();
        }

        // Reopen
        let log = EventLog::open(&dir).unwrap();
        assert_eq!(log.event_count(), 1);

        let mut e2 = make_event(
            "ssn_001",
            EventType::AgentStarted {
                parent_agent_instance_id: None,
            },
        );
        log.append(&mut e2).unwrap();
        assert_eq!(e2.sequence_no, 1);

        let _ = std::fs::remove_dir_all(&dir);
    }

    /// Regression test for #1 in the v0.9.3 Codex adversarial review.
    ///
    /// Multiple `EventLog` instances opened against the same directory must
    /// not collide on `sequence_no`. This simulates what happens when each
    /// `treeship session event` invocation (one per PostToolUse hook firing)
    /// creates a fresh `EventLog` on a shared events.jsonl. Without the
    /// flock-based re-derivation in `append_locked`, every instance sees
    /// the same on-disk count at open time and assigns duplicate sequence
    /// numbers.
    #[cfg(not(target_family = "wasm"))]
    #[test]
    fn concurrent_appends_have_unique_sequence_numbers() {
        use std::sync::Arc;
        use std::thread;

        let dir =
            std::env::temp_dir().join(format!("treeship-evtlog-race-{}", rand::random::<u32>()));
        std::fs::create_dir_all(&dir).unwrap();

        const WRITERS: usize = 16;
        let dir = Arc::new(dir);
        let mut handles = Vec::with_capacity(WRITERS);

        for _ in 0..WRITERS {
            let dir = Arc::clone(&dir);
            handles.push(thread::spawn(move || {
                // Each thread opens its OWN EventLog -- mimics a separate
                // process invocation. Without flock, all threads would see
                // the same line count at open() time.
                let log = EventLog::open(&dir).unwrap();
                let mut e = make_event("ssn_race", EventType::SessionStarted);
                log.append(&mut e).unwrap();
                e.sequence_no
            }));
        }

        let mut seqs: Vec<u64> = handles.into_iter().map(|h| h.join().unwrap()).collect();
        seqs.sort();

        // All sequence numbers must be unique and contiguous 0..WRITERS.
        let expected: Vec<u64> = (0..WRITERS as u64).collect();
        assert_eq!(seqs, expected, "sequence_no collisions under contention");

        // Same invariant from the on-disk file's perspective.
        let log = EventLog::open(&dir).unwrap();
        let read = log.read_all().unwrap();
        assert_eq!(read.len(), WRITERS);
        let mut on_disk: Vec<u64> = read.iter().map(|e| e.sequence_no).collect();
        on_disk.sort();
        assert_eq!(on_disk, expected);

        let _ = std::fs::remove_dir_all(&*dir);
    }

    /// Sidecar lock file must be created mode 0o600 (owner-only) on Unix.
    /// Regression test for #5 in the second Codex adversarial review.
    #[cfg(all(not(target_family = "wasm"), unix))]
    #[test]
    fn lock_file_has_owner_only_permissions() {
        use std::os::unix::fs::PermissionsExt;

        let dir =
            std::env::temp_dir().join(format!("treeship-evtlog-perms-{}", rand::random::<u32>()));
        let log = EventLog::open(&dir).unwrap();

        let mut e = make_event("ssn_perms", EventType::SessionStarted);
        log.append(&mut e).unwrap();

        let lock_path = log.path().with_extension("jsonl.lock");
        let meta = std::fs::metadata(&lock_path).expect("lock file must exist after first append");
        let mode = meta.permissions().mode() & 0o777;
        assert_eq!(
            mode, 0o600,
            "lock file mode is {:o}, expected 0o600 (owner-only)",
            mode
        );

        let _ = std::fs::remove_dir_all(&dir);
    }

    /// A pre-existing lock file (e.g. from a v0.9.2 era crash) with looser
    /// permissions must be tightened to 0o600 on next `EventLog::open`.
    /// Regression test for the third Codex adversarial review.
    #[cfg(all(not(target_family = "wasm"), unix))]
    #[test]
    fn existing_lock_file_is_re_tightened() {
        use std::os::unix::fs::PermissionsExt;

        let dir = std::env::temp_dir().join(format!(
            "treeship-evtlog-retighten-{}",
            rand::random::<u32>()
        ));
        std::fs::create_dir_all(&dir).unwrap();

        // Pre-create a lock file with deliberately loose perms, simulating
        // an upgrade from a CLI version that didn't set 0o600.
        let lock_path = dir.join("events.jsonl.lock");
        std::fs::write(&lock_path, b"").unwrap();
        std::fs::set_permissions(&lock_path, std::fs::Permissions::from_mode(0o644)).unwrap();
        let pre_mode = std::fs::metadata(&lock_path).unwrap().permissions().mode() & 0o777;
        assert_eq!(
            pre_mode, 0o644,
            "test setup: pre-existing perms should be 0o644"
        );

        // First append after upgrade -- should re-tighten.
        let log = EventLog::open(&dir).unwrap();
        let mut e = make_event("ssn_retighten", EventType::SessionStarted);
        log.append(&mut e).unwrap();

        let post_mode = std::fs::metadata(&lock_path).unwrap().permissions().mode() & 0o777;
        assert_eq!(
            post_mode, 0o600,
            "lock file should be re-tightened to 0o600 after open; got {:o}",
            post_mode
        );

        let _ = std::fs::remove_dir_all(&dir);
    }

    /// Counter sidecar must exist after the first append and contain
    /// (count=1, byte_size=size of events.jsonl). This is the happy path
    /// that lets every subsequent append skip the O(N) rescan.
    #[cfg(not(target_family = "wasm"))]
    #[test]
    fn counter_sidecar_written_after_append() {
        let dir =
            std::env::temp_dir().join(format!("treeship-evtlog-counter-{}", rand::random::<u32>()));
        let log = EventLog::open(&dir).unwrap();

        let mut e = make_event("ssn_counter", EventType::SessionStarted);
        log.append(&mut e).unwrap();

        let counter = log.path().with_extension("jsonl.count");
        let bytes = std::fs::read(&counter).expect("counter sidecar must exist after append");
        assert_eq!(bytes.len(), 16, "counter sidecar must be 16 bytes");

        let count = u64::from_le_bytes(bytes[0..8].try_into().unwrap());
        let recorded_size = u64::from_le_bytes(bytes[8..16].try_into().unwrap());
        let actual_size = std::fs::metadata(log.path()).unwrap().len();
        assert_eq!(count, 1, "counter must reflect the one appended event");
        assert_eq!(
            recorded_size, actual_size,
            "counter byte_size ({}) must match events.jsonl size ({})",
            recorded_size, actual_size
        );

        let _ = std::fs::remove_dir_all(&dir);
    }

    /// A missing counter sidecar (fresh install, deleted by user, etc.)
    /// must not break sequence_no assignment. The next append falls back
    /// to an O(N) recount and rewrites the counter.
    #[cfg(not(target_family = "wasm"))]
    #[test]
    fn counter_sidecar_recovers_when_missing() {
        let dir = std::env::temp_dir().join(format!(
            "treeship-evtlog-missing-counter-{}",
            rand::random::<u32>()
        ));

        // Append two events, then nuke the counter sidecar.
        {
            let log = EventLog::open(&dir).unwrap();
            let mut e1 = make_event("ssn_x", EventType::SessionStarted);
            let mut e2 = make_event(
                "ssn_x",
                EventType::AgentStarted {
                    parent_agent_instance_id: None,
                },
            );
            log.append(&mut e1).unwrap();
            log.append(&mut e2).unwrap();
        }
        let counter = dir.join("events.jsonl.count");
        std::fs::remove_file(&counter).expect("counter must exist before deletion");

        // Reopen + append. The third event must get sequence_no=2 even
        // though the counter sidecar is gone.
        let log = EventLog::open(&dir).unwrap();
        assert_eq!(
            log.event_count(),
            2,
            "open() must recount when counter is missing"
        );

        let mut e3 = make_event(
            "ssn_x",
            EventType::SessionClosed {
                summary: None,
                duration_ms: None,
            },
        );
        log.append(&mut e3).unwrap();
        assert_eq!(e3.sequence_no, 2);
        assert!(counter.exists(), "counter must be rewritten after recount");

        let _ = std::fs::remove_dir_all(&dir);
    }

    /// A short-read or garbage counter sidecar (corrupted, partial write,
    /// truncated by external tool) must not be trusted. The size mismatch
    /// path covers the "wrong content" case for a 16-byte file too.
    #[cfg(not(target_family = "wasm"))]
    #[test]
    fn counter_sidecar_recovers_when_corrupt() {
        let dir = std::env::temp_dir().join(format!(
            "treeship-evtlog-corrupt-counter-{}",
            rand::random::<u32>()
        ));

        {
            let log = EventLog::open(&dir).unwrap();
            let mut e = make_event("ssn_corrupt", EventType::SessionStarted);
            log.append(&mut e).unwrap();
        }
        // Truncate the counter to a non-16 length.
        let counter = dir.join("events.jsonl.count");
        std::fs::write(&counter, b"junk").unwrap();

        let log = EventLog::open(&dir).unwrap();
        assert_eq!(
            log.event_count(),
            1,
            "short-read counter must be ignored, recount kicks in"
        );

        let _ = std::fs::remove_dir_all(&dir);
    }

    /// A counter that recorded the wrong byte_size (someone or something
    /// appended to events.jsonl behind our back) must not be trusted.
    /// This is the crash-recovery path: peer wrote events.jsonl but
    /// crashed before fsyncing the counter, so the recorded size is stale.
    #[cfg(not(target_family = "wasm"))]
    #[test]
    fn counter_sidecar_recovers_when_size_disagrees() {
        let dir = std::env::temp_dir().join(format!(
            "treeship-evtlog-stale-counter-{}",
            rand::random::<u32>()
        ));

        {
            let log = EventLog::open(&dir).unwrap();
            let mut e = make_event("ssn_stale", EventType::SessionStarted);
            log.append(&mut e).unwrap();
        }

        // Simulate a crash mid-append: append one extra raw line to
        // events.jsonl WITHOUT updating the counter. Now the counter
        // says (1, S) but events.jsonl is (S + |line|) bytes.
        let events_path = dir.join("events.jsonl");
        let mut extra = make_event(
            "ssn_stale",
            EventType::AgentStarted {
                parent_agent_instance_id: None,
            },
        );
        extra.sequence_no = 999; // intentionally wrong; will be overwritten on read
        let mut line = serde_json::to_vec(&extra).unwrap();
        line.push(b'\n');
        let mut f = std::fs::OpenOptions::new()
            .append(true)
            .open(&events_path)
            .unwrap();
        std::io::Write::write_all(&mut f, &line).unwrap();
        std::io::Write::flush(&mut f).unwrap();

        // Re-open. The size mismatch must trigger a recount; we should see 2.
        let log = EventLog::open(&dir).unwrap();
        assert_eq!(
            log.event_count(),
            2,
            "size mismatch must force recount, ignoring stale counter"
        );

        let _ = std::fs::remove_dir_all(&dir);
    }

    /// The counter sidecar fix must not break the cross-process race
    /// safety established by the flock layer. This is the same shape as
    /// `concurrent_appends_have_unique_sequence_numbers` but exists to
    /// guard against a regression where the counter is read OUTSIDE the
    /// lock, which would let two writers both see count=N and assign N
    /// to two different events.
    #[cfg(not(target_family = "wasm"))]
    #[test]
    fn counter_sidecar_preserves_concurrent_uniqueness() {
        use std::sync::Arc;
        use std::thread;

        let dir = std::env::temp_dir().join(format!(
            "treeship-evtlog-counter-race-{}",
            rand::random::<u32>()
        ));
        std::fs::create_dir_all(&dir).unwrap();

        const WRITERS: usize = 16;
        let dir = Arc::new(dir);
        let mut handles = Vec::with_capacity(WRITERS);

        for _ in 0..WRITERS {
            let dir = Arc::clone(&dir);
            handles.push(thread::spawn(move || {
                let log = EventLog::open(&dir).unwrap();
                let mut e = make_event("ssn_counter_race", EventType::SessionStarted);
                log.append(&mut e).unwrap();
                e.sequence_no
            }));
        }

        let mut seqs: Vec<u64> = handles.into_iter().map(|h| h.join().unwrap()).collect();
        seqs.sort();
        let expected: Vec<u64> = (0..WRITERS as u64).collect();
        assert_eq!(
            seqs, expected,
            "counter must not bypass the flock race protection"
        );

        // Counter should reflect the final state.
        let log = EventLog::open(&dir).unwrap();
        assert_eq!(log.event_count(), WRITERS as u64);

        let _ = std::fs::remove_dir_all(&*dir);
    }

    /// Counter sidecar must be created mode 0o600 (owner-only) on Unix --
    /// same scoping as the lock file; the existence of a counter is a
    /// session signal that doesn't need to leak to other users.
    #[cfg(all(not(target_family = "wasm"), unix))]
    #[test]
    fn counter_sidecar_has_owner_only_permissions() {
        use std::os::unix::fs::PermissionsExt;

        let dir = std::env::temp_dir().join(format!(
            "treeship-evtlog-counter-perms-{}",
            rand::random::<u32>()
        ));
        let log = EventLog::open(&dir).unwrap();

        let mut e = make_event("ssn_counter_perms", EventType::SessionStarted);
        log.append(&mut e).unwrap();

        let counter = log.path().with_extension("jsonl.count");
        let mode = std::fs::metadata(&counter).unwrap().permissions().mode() & 0o777;
        assert_eq!(
            mode, 0o600,
            "counter sidecar mode is {:o}, expected 0o600 (owner-only)",
            mode
        );

        let _ = std::fs::remove_dir_all(&dir);
    }

    /// P0 regression (audit lane F): under heavy hook contention, multiple
    /// writers used to fall through and append without the flock when the
    /// 500ms poll exhausted, producing duplicate `sequence_no` values. The
    /// blocking `lock_exclusive` fix means every writer must hold the
    /// flock across both the counter read and the event write.
    ///
    /// This test spawns 8 threads, each calling `append` 25 times on its
    /// own `EventLog` (mimicking 8 separate hook processes each appending
    /// a burst of events). After the join, the on-disk log must contain
    /// exactly 8*25=200 events with `sequence_no` exactly the contiguous
    /// range 0..200, no duplicates and no gaps.
    #[cfg(not(target_family = "wasm"))]
    #[test]
    fn p0_no_duplicate_sequence_under_burst_contention() {
        use std::sync::Arc;
        use std::thread;

        const THREADS: usize = 8;
        const PER_THREAD: usize = 25;
        const EXPECTED: usize = THREADS * PER_THREAD;

        let dir = std::env::temp_dir().join(format!(
            "treeship-evtlog-p0-burst-{}",
            rand::random::<u32>()
        ));
        std::fs::create_dir_all(&dir).unwrap();
        let dir = Arc::new(dir);

        let mut handles = Vec::with_capacity(THREADS);
        for t in 0..THREADS {
            let dir = Arc::clone(&dir);
            handles.push(thread::spawn(move || -> Vec<u64> {
                // Each thread opens its own EventLog -- this is the
                // per-process model the audit flagged: every PostToolUse
                // invocation is a fresh handle on the shared log.
                let log = EventLog::open(&dir).unwrap();
                let mut seen = Vec::with_capacity(PER_THREAD);
                for i in 0..PER_THREAD {
                    let mut e =
                        make_event(&format!("ssn_burst_{}_{}", t, i), EventType::SessionStarted);
                    log.append(&mut e).unwrap();
                    seen.push(e.sequence_no);
                }
                seen
            }));
        }

        // Collect what each thread saw locally.
        let mut all_returned: Vec<u64> = handles
            .into_iter()
            .flat_map(|h| h.join().unwrap())
            .collect();
        all_returned.sort();

        let expected: Vec<u64> = (0..EXPECTED as u64).collect();
        assert_eq!(
            all_returned, expected,
            "returned sequence_no values must be a contiguous range 0..{} \
             with no duplicates and no gaps",
            EXPECTED
        );

        // Truth source: the on-disk log itself. Verify (a) count and
        // (b) sequence_no is exactly 0..EXPECTED on the persisted events.
        let log = EventLog::open(&dir).unwrap();
        let events = log.read_all().unwrap();
        assert_eq!(
            events.len(),
            EXPECTED,
            "on-disk event count must be exactly {} (got {})",
            EXPECTED,
            events.len()
        );
        let mut on_disk: Vec<u64> = events.iter().map(|e| e.sequence_no).collect();
        on_disk.sort();
        assert_eq!(
            on_disk, expected,
            "on-disk sequence_no must be a contiguous range with no duplicates and no gaps"
        );

        // Counter sidecar must agree with both.
        assert_eq!(log.event_count(), EXPECTED as u64);

        let _ = std::fs::remove_dir_all(&*dir);
    }

    /// Companion stress test for the lock-file lifecycle. Each append
    /// opens the lock file, locks it, writes, unlocks, and drops the FD.
    /// Repeatedly creating + dropping `EventLog`s in a tight loop must
    /// not panic, must not leak FDs we can detect (no `EMFILE` after
    /// hundreds of iterations on a default ulimit), and must produce a
    /// log with contiguous sequence numbers.
    #[cfg(not(target_family = "wasm"))]
    #[test]
    fn lock_file_handles_drop_cleanly_under_churn() {
        let dir = std::env::temp_dir().join(format!(
            "treeship-evtlog-fd-churn-{}",
            rand::random::<u32>()
        ));
        std::fs::create_dir_all(&dir).unwrap();

        // 500 sequential open + append + drop cycles. Far below the
        // default macOS/Linux soft limit (256/1024) for a sustained
        // leak, but plenty to catch one-per-iteration FD leaks.
        const ITERS: usize = 500;
        for i in 0..ITERS {
            let log = EventLog::open(&dir).unwrap();
            let mut e = make_event(&format!("ssn_churn_{}", i), EventType::SessionStarted);
            log.append(&mut e).unwrap();
            // log drops here -> lock_file FD already closed inside append.
        }

        let log = EventLog::open(&dir).unwrap();
        let events = log.read_all().unwrap();
        assert_eq!(events.len(), ITERS);
        let mut seqs: Vec<u64> = events.iter().map(|e| e.sequence_no).collect();
        seqs.sort();
        let expected: Vec<u64> = (0..ITERS as u64).collect();
        assert_eq!(
            seqs, expected,
            "no FD leak should still produce contiguous seqs"
        );

        let _ = std::fs::remove_dir_all(&dir);
    }
}

#[cfg(test)]
mod open_race_tests {
    use super::tests::make_event;
    use super::*;
    use crate::session::EventType;
    use std::sync::Arc;
    use std::thread;

    /// Reproduction harness for #275.
    ///
    /// The burst test opens one `EventLog` per thread, which is the real
    /// model: every hook invocation is a fresh handle on a shared log. That
    /// makes `open()` part of the contended path, and `open()` performs an
    /// unlocked read-modify-write of the counter sidecar
    /// (`read_counter_or_recount` rewrites it on a stale/missing counter)
    /// while another process may hold the append lock.
    ///
    /// More threads and more rounds than the original: the failure is timing
    /// dependent and did not reproduce locally at 8x25 across 20 runs.
    #[test]
    fn open_and_append_interleaved_keep_sequences_unique() {
        const THREADS: usize = 16;
        const PER_THREAD: usize = 12;
        const ROUNDS: usize = 12;

        for round in 0..ROUNDS {
            let dir =
                std::env::temp_dir().join(format!("ts-open-race-{}-{}", std::process::id(), round));
            let _ = std::fs::remove_dir_all(&dir);
            std::fs::create_dir_all(&dir).unwrap();
            let dir = Arc::new(dir);

            let handles: Vec<_> = (0..THREADS)
                .map(|t| {
                    let dir = Arc::clone(&dir);
                    thread::spawn(move || -> Vec<u64> {
                        let mut seen = Vec::with_capacity(PER_THREAD);
                        for i in 0..PER_THREAD {
                            // Re-open every iteration. This is the point: it
                            // puts open() inside the contention window
                            // instead of once before it.
                            let log = EventLog::open(&dir).unwrap();
                            let mut e =
                                make_event(&format!("ssn_race_{t}_{i}"), EventType::SessionStarted);
                            log.append(&mut e).unwrap();
                            seen.push(e.sequence_no);
                        }
                        seen
                    })
                })
                .collect();

            let mut all: Vec<u64> = handles
                .into_iter()
                .flat_map(|h| h.join().unwrap())
                .collect();
            all.sort_unstable();
            let expected: Vec<u64> = (0..(THREADS * PER_THREAD) as u64).collect();
            assert_eq!(
                all, expected,
                "round {round}: sequence_no values must be a contiguous range \
                 with no duplicates and no gaps"
            );
            let _ = std::fs::remove_dir_all(&*dir);
        }
    }
}