chokidar stats are captured at poll time and may not reflect the file's
final byte length by the time vault.read() is called. The downstream
integrity check compares stat.size to content length; a mismatch causes
other LiveSync clients to reject the file as corrupted.
Fix by updating file.stat.size from the actual content in read() and
readBinary().
Co-authored-by: Joysimple <Joysimple@users.noreply.github.com>
- Add chokidar ^4.0.0 as dependency (root package.json, runtime-package.json)
- Mark chokidar as external in vite.config.ts (not bundled, loaded at runtime)
- Implement CLIWatchAdapter.beginWatch() with chokidar:
- ignoreInitial: true (startup files handled by mirror scan)
- awaitWriteFinish to prevent partial-write events
- Excludes dotfiles and .livesync/ directory at watcher level
- Maps add/change/unlink/addDir/unlinkDir to IStorageEventWatchHandlers
- Fatal error handler: logs clearly and releases watcher resources
- Add close() to CLIWatchAdapter, StorageEventManagerCLI for clean shutdown
- Register onUnload hook in CLIServiceModules to close watcher on shutdown
- Add daemon command to help text and --interval/-i flag for polling mode
- Capture original sync settings before suspendAllSync() clobbers them
- Implement daemon startup: mirror scan → restore settings → applySettings()
which triggers the full suspend/resume lifecycle and starts the _changes feed
- Guard processSynchroniseResult no-op to non-daemon commands so default
handler writes incoming CouchDB changes to the local filesystem
- Polling mode: restore settings + clearInterval-safe try/catch error handling
- Warn when both liveSync and syncOnStart are false after restore (no-op config)
- Fix: only block indefinitely if daemon startup succeeded
On Linux, fs.Stats.mtimeMs and ctimeMs return floats with sub-millisecond
precision derived from the kernel's nanosecond filesystem mtime. Stored
raw, this produces document timestamps like 1778511180024.462 in CouchDB
rather than integer milliseconds.
Mobile clients running LiveSync 0.25.60 have been observed to crash when
processing change-feed updates carrying non-integer millisecond timestamps
from CLI-written documents. Desktop and mobile GUI plugins write integer
milliseconds, so the crash only manifests when the headless CLI on Linux
is the source. Whether the issue was introduced in 0.25.60 or had been
latent in earlier versions hasn't been investigated; 0.25.60 is the
version where the crash was confirmed and the fix verified.
Floor the values at every stat-read site (six across three adapters and
one command) so CLI-written documents carry integer-millisecond
timestamps consistent with the rest of the mesh.
