przn 0.5.0 → 0.6.0

data/lib/przn/controller.rb

@@ -2,73 +2,389 @@
 
 module Przn
   class Controller
-    def initialize(presentation, terminal, renderer, audience_link: nil)
+    # `source_file:` / `theme_path:` enable the `r` key to re-read the
+    # deck (and the theme.yml, if one is in play) from disk so authors
+    # can iterate on the markdown without restarting the session.
+    # `theme_path` is the explicit `--theme PATH` value; when nil, the
+    # auto-discover path is re-run on each reload (so dropping a
+    # `theme.yml` next to the deck mid-session also takes effect).
+    def initialize(presentation, terminal, renderer, audience_link: nil,
+                   source_file: nil, theme_path: nil, watch: true)
       @presentation = presentation
       @terminal = terminal
       @renderer = renderer
       @audience_link = audience_link
+      @source_file = source_file
+      @theme_path = theme_path
+      @watch_enabled = watch && !@source_file.nil?
       @preload_gen = 0
+      # Within-slide step counter for incremental reveals (one
+      # bump per `<wait/>`). Space walks these before flipping to
+      # the next slide; arrow-left walks back into the previous
+      # slide's last step. Always resets to 0 on slide jumps and
+      # reloads.
+      @slide_step = 0
     end
 
+    # Seconds between runner-bar refresh ticks. The 🐢 only crawls
+    # across the track over the full talk's duration, so once a second
+    # is plenty for visible movement without being noisy.
+    RUNNER_BAR_TICK_SECONDS = 1.0
+
+    # One decoded SGR mouse report. `button` is 0 (left) / 1 (middle) /
+    # 2 (right) / 64+ (wheel); `kind` is :press, :release or :drag
+    # (motion with a button held). `col` / `row` are 1-based cells.
+    Mouse = Struct.new(:button, :col, :row, :kind)
+
+    # An in-progress 🐇 drag: where on the glyph it was grabbed, and the
+    # slide index the rabbit currently sits at.
+    RunnerDrag = Struct.new(:offset, :target)
+
     def run
       @started_at = Time.now
       @terminal.enter_alt_screen
       @terminal.hide_cursor
+      @terminal.enable_mouse
+      EchoesClient.hide_pointer(io_out: @terminal)
+      warm_code_highlight_cache
       render_current
+      start_runner_bar_thread
+      start_file_watcher
 
       @terminal.raw do
         loop do
-          case read_key
+          case (event = read_event)
+          when Mouse
+            handle_mouse(event)
           when :right, :down, 'l', 'j', ' '
-            @presentation.next_slide
-            render_current
+            advance_step_or_slide
           when :left, :up, 'h', 'k'
-            @presentation.prev_slide
-            render_current
+            retreat_step_or_slide
           when 'g'
             @presentation.first_slide!
+            @slide_step = 0
             render_current
           when 'G'
             @presentation.last_slide!
+            @slide_step = 0
             render_current
+          when 'r'
+            reload
           when 'q', "\x03"
             break
           end
         end
       end
     ensure
+      stop_file_watcher
+      stop_runner_bar_thread
       @preload_gen += 1
       @preload_thread&.join
+      @warmup_thread&.kill
+      @warmup_thread&.join
       if @audience_link
-        AudienceLink.send(@audience_link, type: "quit")
+        AudienceLink.send_message(@audience_link, type: "quit")
         @audience_link.close
       end
       @terminal.write "\e]7772;bg-clear\a"
       @terminal.write ImageUtil.kitty_clear_all if ImageUtil.kitty_terminal?
       @terminal.show_cursor
+      EchoesClient.show_pointer(io_out: @terminal)
       @terminal.leave_alt_screen
     end
 
     private
 
+    # Re-read the source markdown (and theme.yml, if any) from disk
+    # and re-render. Any IO / parse error is swallowed so a bad save
+    # mid-edit doesn't kill the running session — the user just sees
+    # the previous slide until the next reload succeeds.
+    #
+    # `jump_to_change:` lets the auto-watcher request a jump to the
+    # first slide whose source markdown changed (so a save lands the
+    # presenter on the slide they just edited). The manual `r`
+    # binding keeps the historical behaviour: stay on the current
+    # slide index, clamped to the new deck's bounds.
+    def reload(jump_to_change: false)
+      return unless @source_file
+      markdown = File.read(@source_file)
+      new_presentation = Parser.parse(markdown)
+      return if new_presentation.total.zero?
+
+      new_theme =
+        if @theme_path
+          Theme.load(@theme_path)
+        else
+          Theme.auto_discover(near: @source_file)
+        end
+      # auto_discover returns nil when no theme.yml sits next to the
+      # deck; fall back to Theme.default so the renderer never has a
+      # nil theme (matching what Renderer#initialize does on first
+      # construction).
+      new_theme ||= Theme.default
+
+      target = jump_to_change ? first_changed_slide(@last_source, markdown) : nil
+      target ||= [@presentation.current, new_presentation.total - 1].min
+      new_presentation.go_to([target, new_presentation.total - 1].min)
+      @presentation = new_presentation
+      @last_source = markdown
+      @renderer.theme = new_theme
+      # Refresh the renderer's deck-wide ref index — without this an
+      # `r` reload would leave <ref id="..."/> resolving against the
+      # old presentation's blocks.
+      @renderer.presentation = new_presentation
+      @slide_step = 0
+      render_current
+    rescue StandardError
+      # Don't crash the session on a bad edit; keep showing the old deck.
+    end
+
+    # Split the old and new source markdown on `# ` h1 boundaries and
+    # return the index of the first chunk that differs — i.e., the
+    # slide the author most likely just edited. Falls back to nil
+    # when nothing differs (no jump) or when there's no prior snapshot
+    # yet (first reload after start — `r` semantics).
+    def first_changed_slide(old_src, new_src)
+      return nil unless old_src
+      old_chunks = split_into_slide_chunks(old_src).map { |c| c.rstrip }
+      new_chunks = split_into_slide_chunks(new_src).map { |c| c.rstrip }
+      idx = (0...[old_chunks.size, new_chunks.size].max).find { |i|
+        old_chunks[i] != new_chunks[i]
+      }
+      idx
+    end
+
+    def split_into_slide_chunks(src)
+      chunks = []
+      current = +''
+      src.each_line do |line|
+        if line.start_with?('# ') && !current.empty?
+          chunks << current
+          current = +''
+        end
+        current << line
+      end
+      chunks << current unless current.empty?
+      chunks
+    end
+
+    # Background polling watcher over the source markdown and
+    # (optionally) the deck's theme.yml. On change it pushes one byte
+    # into `@reload_pipe_w` to wake `read_event`'s select loop; the
+    # main thread then runs `reload` synchronously. Idempotent — a
+    # burst of writes (some editors save atomically with multiple
+    # mtime ticks) collapses into one reload via select's "ready
+    # again" semantics.
+    def start_file_watcher
+      return unless @watch_enabled
+      @last_source = File.read(@source_file) rescue nil
+      @reload_pipe_r, @reload_pipe_w = IO.pipe
+      paths = [@source_file]
+      paths << @theme_path if @theme_path
+      paths << File.join(File.dirname(File.expand_path(@source_file)), 'theme.yml') unless @theme_path
+      @file_watcher = FileWatcher.new(paths)
+      @file_watcher.start do |changed|
+        # Source-file change → request a jump-to-edited-slide; theme
+        # / sidecar changes only re-render in place.
+        tag = changed == @source_file ? 's' : 't'
+        begin
+          @reload_pipe_w.write_nonblock(tag)
+        rescue IO::WaitWritable, Errno::EPIPE
+          # Pipe full or closed during teardown — drop the signal;
+          # the next save will write its own.
+        end
+      end
+    end
+
+    def stop_file_watcher
+      @file_watcher&.stop
+      @reload_pipe_w&.close
+      @reload_pipe_r&.close
+    rescue StandardError
+    end
+
     def render_current
       @renderer.render(
         @presentation.current_slide,
         current: @presentation.current,
         total: @presentation.total,
-        started_at: @started_at
+        started_at: @started_at,
+        step: @slide_step
       )
       if @audience_link
-        AudienceLink.send(@audience_link,
-                          type: "goto",
-                          index: @presentation.current,
-                          started_at: @started_at.to_f)
+        AudienceLink.send_message(@audience_link,
+                                  type: "goto",
+                                  index: @presentation.current,
+                                  started_at: @started_at.to_f)
       end
       schedule_preload
     end
 
+    # Space / right / down: reveal the next step on the current slide
+    # if there's one left; otherwise jump to the next slide (and reset
+    # the step counter to 0 so the new slide starts collapsed).
+    #
+    # When the just-revealed step contains an animated `<action
+    # duration="...">`, run a synchronous animation loop that emits
+    # `progress` frames at 30fps before settling on the final
+    # `render_current`. The loop blocks the key handler — additional
+    # key presses queue and process after the animation finishes.
+    def advance_step_or_slide
+      total = @renderer.step_count(@presentation.current_slide)
+      if @slide_step < total - 1
+        @slide_step += 1
+        duration_ms = @renderer.max_duration_for_step(@presentation.current_slide, @slide_step)
+        if duration_ms > 0
+          animate_step(duration_ms)
+        else
+          render_current
+        end
+      elsif @presentation.last_slide?
+        # Nothing further to advance to — keep the slide rendered as-is
+        # rather than running a fade-out that would land on the same slide.
+        render_current
+      else
+        cross_slide_boundary { @presentation.next_slide; @slide_step = 0 }
+      end
+    end
+
+    # Default frame rate for the animation loop. 30fps is plenty for
+    # the kinds of single-element moves przn does — and on a busy
+    # slide every frame triggers a full slide re-render plus an SVG
+    # re-upload for any moving shape, so doubling to 60fps would
+    # double Echoes' upload churn without obvious benefit.
+    ANIMATION_FPS = 30.0
+
+    # Monotonic-clock animation loop. Emits intermediate
+    # `Renderer#render(progress:)` calls until `progress` hits 1.0,
+    # then calls `render_current` once more so the post-render
+    # audience_link push and `schedule_preload` (skipped per-frame)
+    # run at the canonical state.
+    def animate_step(duration_ms)
+      frame_ms = 1000.0 / ANIMATION_FPS
+      slide = @presentation.current_slide
+      current = @presentation.current
+      total = @presentation.total
+      started = monotonic_ms
+      loop do
+        elapsed = monotonic_ms - started
+        progress = (elapsed / duration_ms).clamp(0.0, 1.0)
+        @renderer.render(
+          slide,
+          current: current,
+          total: total,
+          started_at: @started_at,
+          step: @slide_step,
+          progress: progress
+        )
+        break if progress >= 1.0
+        sleep_for = (frame_ms - (monotonic_ms - started - elapsed)) / 1000.0
+        sleep(sleep_for) if sleep_for > 0
+      end
+      render_current
+    end
+
+    def monotonic_ms
+      Process.clock_gettime(Process::CLOCK_MONOTONIC, :millisecond)
+    end
+
+    # Wrap a slide change with optional fade-out / fade-in halves so a
+    # slide opted into the fade transition via `# Title {transition=fade}`
+    # animates instead of snapping. The two halves are independent:
+    # leaving slide A fades out iff A opted in; arriving at slide B fades
+    # in iff B opted in. Block yields the slide-change mutation (next or
+    # prev) sandwiched between the halves.
+    #
+    # Gated on Echoes — non-Echoes terminals skip the fade entirely and
+    # just snap, so the OSC 7772 cell-alpha bytes don't reach a terminal
+    # that would drop them and leave the loop's per-frame re-renders
+    # visible as flicker.
+    def cross_slide_boundary
+      outgoing = @presentation.current_slide
+      duration = slide_fade_duration_ms
+      can_fade = ImageUtil.echoes_terminal? && duration > 0
+      out_fade = can_fade && fade_slide?(outgoing)
+
+      animate_slide_fade(outgoing, from: 1.0, to: 0.0, duration_ms: duration) if out_fade
+      yield
+      incoming = @presentation.current_slide
+      in_fade  = can_fade && fade_slide?(incoming)
+      animate_slide_fade(incoming, from: 0.0, to: 1.0, duration_ms: duration) if in_fade
+      render_current
+    end
+
+    # `{transition=fade}` on the slide's h1 IAL → @attrs[:transition] == "fade".
+    # Symbols and strings both accepted for forward-compat.
+    def fade_slide?(slide)
+      v = slide&.attrs && slide.attrs[:transition]
+      v.to_s == 'fade'
+    end
+
+    def slide_fade_duration_ms
+      theme = @renderer.respond_to?(:theme) ? @renderer.theme : nil
+      ((theme && theme.transition && theme.transition[:duration]) || 200).to_i
+    end
+
+    # Per-half slide fade — same shape as `animate_step` but lerps a
+    # slide-level `opacity:` from `from` → `to` over `duration_ms`.
+    def animate_slide_fade(slide, from:, to:, duration_ms:)
+      frame_ms = 1000.0 / ANIMATION_FPS
+      current = @presentation.current
+      total = @presentation.total
+      started = monotonic_ms
+      loop do
+        elapsed = monotonic_ms - started
+        progress = (elapsed / duration_ms.to_f).clamp(0.0, 1.0)
+        alpha = from + (to - from) * progress
+        @renderer.render(
+          slide,
+          current: current,
+          total: total,
+          started_at: @started_at,
+          step: @slide_step,
+          opacity: alpha
+        )
+        break if progress >= 1.0
+        sleep_for = (frame_ms - (monotonic_ms - started - elapsed)) / 1000.0
+        sleep(sleep_for) if sleep_for > 0
+      end
+    end
+
+    # Left / up: step backward within the current slide if there's
+    # an earlier step to fall back to; otherwise jump to the previous
+    # slide and LAND on its last step so backtracking shows the full
+    # slide instead of its initial-collapsed view. Mirrors how Keynote
+    # walks builds in reverse.
+    def retreat_step_or_slide
+      if @slide_step > 0
+        @slide_step -= 1
+        render_current
+      else
+        prev_idx = @presentation.current - 1
+        if prev_idx < 0
+          # Already on slide 0; stay put. No fade.
+          render_current
+        else
+          cross_slide_boundary do
+            @presentation.prev_slide
+            @slide_step = [@renderer.step_count(@presentation.current_slide) - 1, 0].max
+          end
+        end
+      end
+    end
+
+    # Number of slides on each side of the current slide to pre-upload
+    # in the background. Higher = snappier multi-step navigation but
+    # more startup work per slide change (and more memory tied up in
+    # Echoes' image cache). 3 covers the "I'm about to skip a couple
+    # of slides ahead" case without crowding out the LRU when shape-
+    # heavy decks have lots of image_ids per slide.
+    PRELOAD_RADIUS = 3
+
     # Kick off a background thread that pre-uploads images for the slides the
-    # user is most likely to visit next (the immediate neighbors). Uses a
+    # user is most likely to visit next. Walks PRELOAD_RADIUS slides in each
+    # direction, interleaved (next, prev, next+1, prev+1, …) so the most-
+    # likely visit gets warmed first if the user keys quickly. Uses a
     # generation counter so a navigation that lands while a preload is still
     # running causes that preload to exit early instead of stacking work.
     def schedule_preload
@@ -76,7 +392,8 @@ module Przn
       gen = @preload_gen
       cur = @presentation.current
       total = @presentation.total
-      indices = [cur + 1, cur - 1].select { |i| i.between?(0, total - 1) }
+      indices = (1..PRELOAD_RADIUS).flat_map { |d| [cur + d, cur - d] }
+                                    .select { |i| i.between?(0, total - 1) }
       return if indices.empty?
 
       @preload_thread = Thread.new do
@@ -89,6 +406,72 @@ module Przn
       end
     end
 
+    # Fire off a background thread at session start that runs every
+    # fenced code block in the deck through CodeHighlighter so the
+    # `require 'rouge'` + per-lexer autoload + per-block tokenize
+    # costs land *here* instead of on the main render path. By the
+    # time the user navigates anywhere, the cache is already warm.
+    # Slide 0 races the first render — whichever side tokenizes
+    # first writes the cache and the other side finds the hit; no
+    # correctness issue, just an "if it wins, the first visit is
+    # fast too" optimization. Background work, never blocks startup,
+    # silent on failure.
+    def warm_code_highlight_cache
+      @warmup_thread = Thread.new do
+        @presentation.slides.each do |slide|
+          slide.blocks.each do |block|
+            next unless block[:type] == :code_block && block[:language]
+            CodeHighlighter.highlight(block[:content], block[:language])
+          end
+        end
+      rescue StandardError
+        # Background warmup must never crash the session.
+      end
+    end
+
+    # Spawn a thread that calls `redraw_runner_bar` once per second so
+    # the 🐢 turtle visibly crawls across the screen against real time
+    # instead of only updating on slide navigation. No-op when the
+    # theme didn't opt into the runner bar (no `counter.duration`) or
+    # we're in export / audience-only modes.
+    def start_runner_bar_thread
+      return unless runner_bar_thread_eligible?
+      @runner_bar_stop = false
+      @runner_bar_thread = Thread.new do
+        until @runner_bar_stop
+          sleep RUNNER_BAR_TICK_SECONDS
+          break if @runner_bar_stop
+          begin
+            @renderer.redraw_runner_bar(
+              current: @presentation.current,
+              total: @presentation.total,
+              started_at: @started_at
+            )
+          rescue StandardError
+            # Background work must not crash the presentation.
+          end
+        end
+      end
+    end
+
+    def stop_runner_bar_thread
+      @runner_bar_stop = true
+      # Wake the sleep so the thread exits promptly instead of waiting
+      # out the full tick interval on quit.
+      @runner_bar_thread&.wakeup rescue nil
+      @runner_bar_thread&.join
+    end
+
+    def runner_bar_thread_eligible?
+      return false unless @renderer.respond_to?(:redraw_runner_bar)
+      return false if @renderer.respond_to?(:export_mode) && @renderer.export_mode
+      theme = @renderer.respond_to?(:theme) ? @renderer.theme : nil
+      return false unless theme
+      counter = theme.counter
+      duration = counter && counter[:duration]
+      duration && Theme.parse_duration(duration).to_i.positive?
+    end
+
     def read_key
       c = $stdin.getc
       return nil unless c
@@ -102,6 +485,7 @@ module Przn
         when '[D' then :left
         when '[H' then :home
         when '[F' then :end
+        when '[<' then read_mouse
         else :escape
         end
       else
@@ -110,5 +494,112 @@ module Przn
     rescue IO::WaitReadable
       :escape
     end
+
+    # SGR mouse report (`Terminal#enable_mouse`), the `ESC [ <` already
+    # consumed: `Cb ; Cx ; Cy` then `M` on press / drag, `m` on release.
+    # Always returns a Mouse (or :mouse for a garbled report) so the
+    # whole sequence is swallowed rather than leaking digits into the
+    # key loop.
+    def read_mouse
+      params = +''
+      while (c = $stdin.getc)
+        break if c == 'M' || c == 'm'
+        return :mouse if params.size > 16
+        params << c
+      end
+      return :mouse unless c
+
+      cb, col, row = params.split(';').map(&:to_i)
+      kind = if c == 'm' then :release
+             elsif cb & 32 != 0 then :drag
+             else :press
+             end
+      # Strip the shift / meta / ctrl modifier bits (4 / 8 / 16) and the
+      # motion bit (32), so a modified click still counts as a click.
+      Mouse.new(cb & ~0b111100, col, row, kind)
+    end
+
+    # Left-click is next, right-click is previous — except a left press
+    # on the 🐇 of the runner bar, which grabs it: dragging scrubs the
+    # deck, the slide following the rabbit along the track.
+    def handle_mouse(ev)
+      if @runner_drag && ev.kind != :press
+        drag_runner(ev)
+        return
+      end
+      @runner_drag = nil
+      return unless ev.kind == :press
+
+      case ev.button
+      when 0
+        offset = @renderer.respond_to?(:runner_grab) &&
+                 @renderer.runner_grab(ev.row, ev.col, current: @presentation.current, total: @presentation.total)
+        if offset
+          @runner_drag = RunnerDrag.new(offset, @presentation.current)
+        else
+          advance_step_or_slide
+        end
+      when 2
+        retreat_step_or_slide
+      end
+    end
+
+    # A fast drag queues a burst of motion reports; rendering a slide
+    # for each would lag behind the pointer. Track the target on every
+    # report but render only once the queue is drained (the latest
+    # report wins), and settle on release for good measure.
+    def drag_runner(ev)
+      drag = @runner_drag
+      case ev.kind
+      when :drag
+        return unless ev.button == 0
+        drag.target = @renderer.runner_slide_at(ev.col - drag.offset, current: @presentation.current, total: @presentation.total)
+        show_runner_target(drag) unless input_pending?
+      when :release
+        @runner_drag = nil
+        show_runner_target(drag)
+      end
+    end
+
+    def show_runner_target(drag)
+      return if drag.target == @presentation.current
+      @presentation.go_to(drag.target)
+      @slide_step = 0
+      render_current
+    end
+
+    def input_pending?
+      !IO.select([$stdin], nil, nil, 0).nil?
+    end
+
+    # Multiplex stdin and the file-watcher pipe. Without the watcher
+    # wired up, this is a thin pass-through to `read_key` so the
+    # `--no-watch` / no-source-file paths behave exactly as they did
+    # before. With the watcher up, `select` blocks until either a
+    # keystroke or a save lands, and a save returns `:reload`.
+    def read_event
+      return read_key unless @reload_pipe_r
+      readable, = IO.select([$stdin, @reload_pipe_r])
+      if readable.include?(@reload_pipe_r)
+        # Drain whatever the watcher queued so a flurry of saves
+        # collapses to a single reload. The tag tells us whether to
+        # jump to the modified slide (source change) or stay put
+        # (theme change).
+        tags = drain_reload_pipe
+        reload(jump_to_change: tags.include?('s'))
+        nil
+      else
+        read_key
+      end
+    end
+
+    def drain_reload_pipe
+      tags = +''
+      loop do
+        tags << @reload_pipe_r.read_nonblock(64)
+      end
+    rescue IO::WaitReadable, EOFError
+      tags
+    end
   end
 end

data/lib/przn/echoes_client.rb

@@ -29,6 +29,26 @@ module Przn
       nil
     end
 
+    # Tell Echoes to hide the OS mouse pointer over its window — same
+    # effect as the user's cmd-shift-p shortcut, but emitted at the
+    # start of a presentation so the pointer doesn't reappear every
+    # time the presenter brushes the trackpad. Non-Echoes terminals
+    # ignore the OSC.
+    def hide_pointer(io_out: $stdout)
+      io_out.write("#{OSC};hide-pointer#{BEL}")
+      io_out.flush if io_out.respond_to?(:flush)
+      true
+    end
+
+    # Restore the OS mouse pointer hidden by `hide_pointer`. Emitted on
+    # presentation teardown so the user gets their cursor back when they
+    # quit, instead of having to reach for cmd-shift-p.
+    def show_pointer(io_out: $stdout)
+      io_out.write("#{OSC};show-pointer#{BEL}")
+      io_out.flush if io_out.respond_to?(:flush)
+      true
+    end
+
     # Open a new Echoes window on the given display, running `argv` (an
     # Array of strings — argv[0] is the executable). `fullscreen:` is a hint.
     # Returns true if the request was emitted; nothing in the protocol confirms

data/lib/przn/file_watcher.rb

@@ -0,0 +1,69 @@
+# frozen_string_literal: true
+
+module Przn
+  # Polling-based watcher over a small fixed set of files. Single
+  # background thread checks `File.mtime` every `interval` seconds; the
+  # first time a path's mtime moves past the snapshot taken at start,
+  # the block runs with that path as its argument.
+  #
+  # Built around the on-save edit cadence: a human typing in their
+  # editor produces at most one save per few seconds, so a 0.5 s poll
+  # is well below human-perception latency and saves us a gem dep on
+  # rb-fsevent / listen.
+  #
+  # Paths that don't exist when `start` is called (e.g. an optional
+  # `theme.yml` that hasn't been written yet) are watched too — the
+  # mtime starts at nil, so the first save trips the change handler.
+  class FileWatcher
+    DEFAULT_INTERVAL = 0.5
+
+    def initialize(paths, interval: DEFAULT_INTERVAL)
+      @paths    = paths.compact
+      @interval = interval
+      @thread   = nil
+      @stop     = false
+    end
+
+    def start(&on_change)
+      return if @thread
+      mtimes = snapshot
+      @stop = false
+      @thread = Thread.new do
+        Thread.current.abort_on_exception = false
+        until @stop
+          sleep @interval
+          break if @stop
+          @paths.each do |path|
+            current = mtime_for(path)
+            next if current == mtimes[path]
+            mtimes[path] = current
+            begin
+              on_change.call(path)
+            rescue StandardError
+              # Don't kill the watcher because a single notification
+              # raised — the next save still gets a chance.
+            end
+          end
+        end
+      end
+    end
+
+    def stop
+      @stop = true
+      @thread&.join
+      @thread = nil
+    end
+
+    private
+
+    def snapshot
+      @paths.each_with_object({}) { |p, h| h[p] = mtime_for(p) }
+    end
+
+    def mtime_for(path)
+      File.mtime(path)
+    rescue Errno::ENOENT
+      nil
+    end
+  end
+end