fswatch.cr

Cross-platform filesystem watching for Crystal. Port & adapters core with native event backends and a universal polling fallback. Inspired by fswatch / libfswatch by Enrico M. Crisostomo — the event model follows its Created/Removed/Moved/Updated/IsDir flags; this is an independent pure-Crystal implementation with no C dependency.

require "fswatch"

watch = FSWatch.watch("src") do |event|
  if event.created?
    puts "new: #{event.path}"
  elsif event.deleted?
    puts "gone: #{event.path}"
  end
end

# ... later
watch.stop

Backends

The backend is chosen at compile time; FSWatch::Poll can also be driven directly (it satisfies the same Backend interface as the core uses).

Threading model

The observation loop is a Crystal fiber, not an OS thread: the backend fd is registered with the event loop (epoll/kqueue) and the fiber suspends in read/wait_readable, waking only on events. Consequences:

• no foreign threads ever call into Crystal — handlers run on the cooperative scheduler and may touch any Crystal state freely;
• idle cost is zero (no polling, no spinning);
• the host program must run the scheduler the usual Crystal way — a program that never yields (e.g. a blocking C event loop) gets no events until it does.

Crystal's Thread requires -Dpreview_mt and is therefore not used by this shard.

Platform caveats

• inotify (as documented by libfswatch too): not recursive — the backend arms a watch per directory; children born between a directory's creation and the watch being armed are not reported. Network filesystems are not covered; the kernel queue can overflow (delivered as an Overflow event — rescan any caches).
• kqueue (experimental, not yet CI-tested): a directory's NOTE_WRITE carries no child name, so directory events are coarse (Modified on the directory itself); recursion keeps the watch set in step but events stay unnamed. Exact for watched files.
• poll: moves are reported as Deleted + Created; metadata-only changes are not reported; latency = interval:.

Symlinks

Symlinked directories are watched as leaves and never descended — a symlink loop would otherwise grow the walk/watch set forever.

License

MIT. The event-model inspiration, libfswatch, is dual-licensed GPL-3.0/Apache-2.0 by its author.