perf(maestro): buffer TUI output and skip unchanged frames

The TUI repainted on every 33ms tick whether or not anything had
changed — measured at ~725 bytes/s of escape sequences with nothing
attached and nothing happening. Over a remote terminal that is constant
traffic for no visible change.

Output was also unbuffered: `CrosstermBackend::new(std::io::stdout())`
issues one `write!` per cell run, 3551 of them for a 200x50 repaint,
each taking the global stdout lock. A BufWriter collapses a frame into
one flush (measured: 3551 writes -> 2).

Frames are now keyed on the catalog version, the active pane's screen
revision, an input epoch and the terminal size, and skipped when the key
is unchanged. Session cards carry an advancing duration, so a frame
still never stands for longer than a second.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

crates/maestro-tui/src/app.rs

@@ -11,7 +11,6 @@
 //! `dispatch` / the overlay routing.
 
 use std::collections::{HashMap, HashSet};
-use std::io::Stdout;
 use std::path::PathBuf;
 use std::time::Duration;
 
@@ -19,7 +18,6 @@ use chrono::Utc;
 use crossterm::event::{Event as CtEvent, EventStream};
 use futures_util::StreamExt;
 use maestro_protocol::{routes, BranchInfo, CreateSessionRequest, DaemonClient, NameBatch};
-use ratatui::backend::CrosstermBackend;
 use ratatui::Terminal;
 use uuid::Uuid;
 
@@ -89,8 +87,36 @@ pub struct App {
     /// that can move `Focus::Pane` (attach, tab switch, sidebar toggle,
     /// mouse) is covered without each one remembering to report it.
     focused_session: Option<Uuid>,
+    /// Fingerprint of the last frame drawn. The loop compares a freshly
+    /// computed key against this and skips the repaint when they match — an
+    /// idle TUI used to emit a full diff 30 times a second forever, which over
+    /// a remote terminal is constant traffic for no change on screen.
+    last_frame: Option<FrameKey>,
+    /// When that frame went out, for the periodic refresh below.
+    last_drawn_at: std::time::Instant,
+    /// Bumped by every path that mutates `AppState` outside the catalog and
+    /// the attached pane: keys, mouse, draft results, link transitions. Those
+    /// mutations are too diffuse to fingerprint individually, so the loop
+    /// treats "an input was processed" as "assume the frame is stale".
+    epoch: u64,
 }
 
+/// Everything a frame is built from that is cheap to sample every tick.
+/// Equal keys mean an identical frame, so the repaint can be skipped.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+struct FrameKey {
+    catalog: u64,
+    epoch: u64,
+    /// Active pane's session, its screen revision, and its link state.
+    pane: Option<(Uuid, u64, crate::session_task::SessionState)>,
+    area: ratatui::layout::Size,
+}
+
+/// Session cards carry a duration that advances on its own (`docs/SIDEBAR.md`),
+/// so a frame is never allowed to stand for longer than this even when nothing
+/// else changed.
+const CLOCK_REFRESH: Duration = Duration::from_secs(1);
+
 impl std::fmt::Debug for App {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         f.debug_struct("App")
@@ -170,13 +196,16 @@ impl App {
             last_sent_size: HashMap::new(),
             clicks: ClickTracker::default(),
             focused_session: None,
+            last_frame: None,
+            last_drawn_at: std::time::Instant::now(),
+            epoch: 0,
         }
     }
 
     /// Drive the loop until the user quits. Consumes self.
-    pub async fn run(
+    pub async fn run<B: ratatui::backend::Backend>(
         mut self,
-        terminal: &mut Terminal<CrosstermBackend<Stdout>>,
+        terminal: &mut Terminal<B>,
     ) -> Result<(), TuiError> {
         if let Ok(info) = self.daemon.daemon_info().await {
             self.home_dir = info.home_dir;
@@ -184,7 +213,7 @@ impl App {
         let mut events = EventStream::new();
         loop {
             self.sync_focus().await;
-            self.redraw(terminal)?;
+            self.redraw_if_stale(terminal)?;
 
             tokio::select! {
                 Some(Ok(ev)) = events.next() => {
@@ -193,17 +222,22 @@ impl App {
                     }
                 }
                 Some(update) = self.draft_rx.recv() => {
+                    self.epoch = self.epoch.wrapping_add(1);
                     self.apply_draft_update(update).await;
                 }
                 _ = self.catalog.changed() => {
-                    // Render on next loop iteration.
+                    // Picked up by the frame key's catalog version.
                 }
                 _ = self.link.changed() => {
                     // The status indicator and the restart overlay both read
-                    // the link; redraw on the next loop iteration.
+                    // the link; neither is fingerprinted, so mark stale.
+                    self.epoch = self.epoch.wrapping_add(1);
                 }
                 _ = tokio::time::sleep(Duration::from_millis(33)) => {
                     let effects = self.poll_due_probes_at(std::time::Instant::now());
+                    if !effects.is_empty() {
+                        self.epoch = self.epoch.wrapping_add(1);
+                    }
                     self.perform(effects).await;
                 }
             }
@@ -217,6 +251,10 @@ impl App {
             return Ok(false);
         };
         let key = self.clicks.promote(key, std::time::Instant::now());
+        // Any key can move the cursor, open an overlay or edit a text buffer.
+        // Rather than fingerprint each of those, treat a processed input as
+        // proof the frame is stale.
+        self.epoch = self.epoch.wrapping_add(1);
 
         // The daemon is gone: everything a key could do goes through it, so
         // only the quit binding still resolves until it answers again.
@@ -959,9 +997,46 @@ impl App {
         self.focused_session = wanted;
     }
 
-    fn redraw(
+    /// Sample everything a frame depends on. Cheap: two atomic loads, an
+    /// `RwLock` read of a `u64`, and the backend's size.
+    fn frame_key(&self, area: ratatui::layout::Size) -> FrameKey {
+        let pane = self.state.active_session().and_then(|sid| {
+            self.emulators
+                .get(sid)
+                .map(|h| (sid, h.revision(), h.state()))
+        });
+        FrameKey {
+            catalog: self.catalog.version(),
+            epoch: self.epoch,
+            pane,
+            area,
+        }
+    }
+
+    /// Repaint only when the frame key moved, or the clock refresh is due.
+    fn redraw_if_stale<B: ratatui::backend::Backend>(
         &mut self,
-        terminal: &mut Terminal<CrosstermBackend<Stdout>>,
+        terminal: &mut Terminal<B>,
+    ) -> Result<(), TuiError> {
+        // Polled rather than driven off `Event::Resize`: `Terminal::draw`'s own
+        // autoresize only runs on a draw, and a skipped draw must still notice
+        // the terminal changed shape.
+        let area = terminal
+            .size()
+            .map_err(|e| TuiError::Terminal(e.to_string()))?;
+        let key = self.frame_key(area);
+        if self.last_frame == Some(key) && self.last_drawn_at.elapsed() < CLOCK_REFRESH {
+            return Ok(());
+        }
+        self.redraw(terminal)?;
+        self.last_frame = Some(key);
+        self.last_drawn_at = std::time::Instant::now();
+        Ok(())
+    }
+
+    fn redraw<B: ratatui::backend::Backend>(
+        &mut self,
+        terminal: &mut Terminal<B>,
     ) -> Result<(), TuiError> {
         let snap = self.catalog.snapshot();
         let attached: HashSet<Uuid> = self
@@ -1257,6 +1332,100 @@ mod tests {
         )
     }
 
+    /// Counts how many times the loop actually repainted, by watching the
+    /// `TestBackend` buffer's generation via `draw` calls.
+    fn drew(app: &mut App, terminal: &mut Terminal<ratatui::backend::TestBackend>) -> bool {
+        let before = app.last_drawn_at;
+        app.redraw_if_stale(terminal).unwrap();
+        app.last_drawn_at != before
+    }
+
+    #[test]
+    fn unchanged_frame_is_not_repainted() {
+        let mut app = make_app();
+        let mut terminal = Terminal::new(ratatui::backend::TestBackend::new(80, 24)).unwrap();
+
+        assert!(drew(&mut app, &mut terminal), "first frame must paint");
+        assert!(
+            !drew(&mut app, &mut terminal),
+            "nothing changed; the repaint should be skipped"
+        );
+    }
+
+    #[test]
+    fn a_processed_key_makes_the_frame_stale() {
+        let mut app = make_app();
+        let mut terminal = Terminal::new(ratatui::backend::TestBackend::new(80, 24)).unwrap();
+        drew(&mut app, &mut terminal);
+
+        app.epoch = app.epoch.wrapping_add(1);
+        assert!(drew(&mut app, &mut terminal));
+    }
+
+    #[test]
+    fn a_catalog_event_makes_the_frame_stale() {
+        let mut app = make_app();
+        let mut terminal = Terminal::new(ratatui::backend::TestBackend::new(80, 24)).unwrap();
+        drew(&mut app, &mut terminal);
+
+        app.catalog
+            .apply_event(maestro_protocol::CatalogEvent::Snapshot {
+                sessions: Vec::new(),
+            });
+        assert!(drew(&mut app, &mut terminal));
+    }
+
+    #[test]
+    fn a_resize_makes_the_frame_stale() {
+        let mut app = make_app();
+        let mut terminal = Terminal::new(ratatui::backend::TestBackend::new(80, 24)).unwrap();
+        drew(&mut app, &mut terminal);
+
+        terminal.backend_mut().resize(120, 40);
+        assert!(drew(&mut app, &mut terminal));
+    }
+
+    #[test]
+    fn pane_output_makes_the_frame_stale() {
+        use crate::session_task::{SessionHandle, SessionState};
+        use std::sync::atomic::{AtomicU64, AtomicU8};
+        use std::sync::{Arc, Mutex};
+
+        let mut app = make_app();
+        let mut terminal = Terminal::new(ratatui::backend::TestBackend::new(80, 24)).unwrap();
+
+        let sid = Uuid::from_u128(7);
+        let revision = Arc::new(AtomicU64::new(0));
+        app.emulators.insert(
+            sid,
+            SessionHandle {
+                tx: tokio::sync::mpsc::channel(1).0,
+                screen: Arc::new(Mutex::new(vt100::Parser::new(24, 80, 0))),
+                state: Arc::new(AtomicU8::new(SessionState::Live as u8)),
+                revision: revision.clone(),
+            },
+        );
+        app.state.attach_to_focused(sid);
+        drew(&mut app, &mut terminal);
+        assert!(!drew(&mut app, &mut terminal), "idle pane must not repaint");
+
+        revision.fetch_add(1, std::sync::atomic::Ordering::Release);
+        assert!(drew(&mut app, &mut terminal), "new PTY bytes must repaint");
+    }
+
+    #[test]
+    fn a_stale_clock_repaints_even_with_no_events() {
+        let mut app = make_app();
+        let mut terminal = Terminal::new(ratatui::backend::TestBackend::new(80, 24)).unwrap();
+        drew(&mut app, &mut terminal);
+        assert!(!drew(&mut app, &mut terminal));
+
+        // Session cards show an advancing duration, so a frame may not stand
+        // for longer than CLOCK_REFRESH regardless of the key.
+        app.last_drawn_at = std::time::Instant::now() - CLOCK_REFRESH;
+        assert!(drew(&mut app, &mut terminal));
+    }
+
     #[test]
     fn enter_insert_then_leave() {
         let mut app = make_app();
@@ -2091,6 +2260,7 @@ mod tests {
             tx: mpsc::channel(1).0,
             screen: Arc::clone(&screen),
             state: Arc::new(AtomicU8::new(SessionState::Live as u8)),
+            revision: Arc::new(std::sync::atomic::AtomicU64::new(0)),
         };
         let mut app = make_app();
         app.state.active_tab_mut().pane = Some(sid);

crates/maestro-tui/src/catalog.rs

@@ -87,6 +87,16 @@ impl CatalogStore {
     }
 
     /// Cheap O(sessions) read of current state.
+    /// The version counter alone, without cloning the session list. The render
+    /// loop reads this every tick to decide whether the sidebar is stale.
+    pub fn version(&self) -> u64 {
+        self.inner
+            .state
+            .read()
+            .expect("catalog rwlock poisoned")
+            .version
+    }
+
     pub fn snapshot(&self) -> CatalogSnapshot {
         let s = self.inner.state.read().expect("catalog rwlock poisoned");
         let mut sessions: Vec<Session> = s.sessions.values().cloned().collect();

crates/maestro-tui/src/emulators.rs

@@ -168,6 +168,7 @@ mod tests {
             tx: mpsc::channel(1).0,
             screen: Arc::new(Mutex::new(vt100::Parser::new(24, 80, 0))),
             state: Arc::new(AtomicU8::new(SessionState::Live as u8)),
+            revision: Arc::new(std::sync::atomic::AtomicU64::new(0)),
         };
         emulators.insert(id, handle);
         assert!(emulators.get(id).is_some());

crates/maestro-tui/src/lib.rs

@@ -24,6 +24,7 @@ pub mod state;
 pub mod terminal;
 pub mod text_input;
 
+use std::io::BufWriter;
 use std::path::PathBuf;
 use std::time::Duration;
 
@@ -121,13 +122,23 @@ pub async fn run(cfg: RunConfig) -> Result<(), TuiError> {
     let _guard = terminal::TerminalGuard::enter().map_err(|e| match e {
         terminal::TerminalError::Io(io) => TuiError::Terminal(io.to_string()),
     })?;
-    let backend = CrosstermBackend::new(std::io::stdout());
+    // Buffered: ratatui issues one `write!` per cell run — thousands per
+    // frame — and an unbuffered `Stdout` takes the global stdout lock on every
+    // one of them. A single BufWriter collapses a frame into one flush.
+    let backend = CrosstermBackend::new(BufWriter::with_capacity(
+        FRAME_BUFFER_BYTES,
+        std::io::stdout(),
+    ));
     let mut terminal = Terminal::new(backend).map_err(|e| TuiError::Terminal(e.to_string()))?;
 
     let app = App::new(tui_cfg, catalog, daemon, link);
     app.run(&mut terminal).await
 }
 
+/// One full 200x50 repaint is ~5 KB of escape sequences; 64 KB leaves room for
+/// a large terminal without ever flushing mid-frame.
+const FRAME_BUFFER_BYTES: usize = 64 * 1024;
+
 /// Startup probe budget, applied to connect *and* the whole request: the user is
 /// staring at a terminal, so an unreachable daemon must fail in under a second.
 const HEALTH_TIMEOUT: Duration = Duration::from_millis(500);

crates/maestro-tui/src/session_task.rs

@@ -14,7 +14,7 @@
 //!   parser is **not** reset on reconnect — the last-known frame stays
 //!   visible until new bytes arrive, so the user doesn't see a blank.
 
-use std::sync::atomic::{AtomicU8, Ordering};
+use std::sync::atomic::{AtomicU64, AtomicU8, Ordering};
 use std::sync::{Arc, Mutex};
 use std::time::Duration;
 
@@ -56,11 +56,15 @@ pub enum SessionState {
 ///   here.
 /// - `state` — `Arc<AtomicU8>`: live status of the WS, queryable without
 ///   holding the screen lock.
+/// - `revision` — `Arc<AtomicU64>`: bumped every time the parser's screen is
+///   mutated. The render loop compares it against the last drawn frame's so an
+///   idle pane costs a `load` instead of a repaint.
 #[derive(Clone)]
 pub struct SessionHandle {
     pub tx: mpsc::Sender<SessionInput>,
     pub screen: Arc<Mutex<vt100::Parser>>,
     pub(crate) state: Arc<AtomicU8>,
+    pub(crate) revision: Arc<AtomicU64>,
 }
 
 impl std::fmt::Debug for SessionHandle {
@@ -81,6 +85,12 @@ impl SessionHandle {
         }
     }
 
+    /// Monotonic counter of screen mutations. Only equality across two reads
+    /// is meaningful — the loop uses it to decide whether a frame is stale.
+    pub fn revision(&self) -> u64 {
+        self.revision.load(Ordering::Acquire)
+    }
+
     /// Stop the underlying task. Used when the pane closes and the user has
     /// no other tabs pointing at this session.
     pub fn end(&self) {
@@ -109,9 +119,11 @@ impl SessionTask {
         let (tx, mut rx) = mpsc::channel::<SessionInput>(64);
         let parser = Arc::new(Mutex::new(vt100::Parser::new(rows, cols, SCROLLBACK_ROWS)));
         let state = Arc::new(AtomicU8::new(SessionState::Connecting as u8));
+        let revision = Arc::new(AtomicU64::new(0));
 
         let parser_bg = parser.clone();
         let state_bg = state.clone();
+        let revision_bg = revision.clone();
 
         tokio::spawn(async move {
             let url = ws_url;
@@ -151,6 +163,7 @@ impl SessionTask {
                                 }
                                 SessionInput::Resize { cols, rows } => {
                                     parser_bg.lock().expect("vt100 parser mutex").set_size(rows, cols);
+                                    revision_bg.fetch_add(1, Ordering::Release);
                                     let frame = serde_json::to_string(&ClientControl::Resize {
                                         cols,
                                         rows,
@@ -166,6 +179,7 @@ impl SessionTask {
                             match frame {
                                 Some(Ok(Message::Binary(bytes))) => {
                                     parser_bg.lock().expect("vt100 parser mutex").process(&bytes);
+                                    revision_bg.fetch_add(1, Ordering::Release);
                                 }
                                 Some(Ok(Message::Text(text))) => {
                                     // Ping/Pong control frames travel as text JSON. Reply to Ping.
@@ -202,6 +216,7 @@ impl SessionTask {
             tx,
             screen: parser,
             state,
+            revision,
         }
     }
 }
@@ -228,6 +243,7 @@ mod tests {
             tx: mpsc::channel(1).0,
             screen: Arc::new(Mutex::new(vt100::Parser::new(24, 80, 0))),
             state: Arc::new(AtomicU8::new(SessionState::Connecting as u8)),
+            revision: Arc::new(AtomicU64::new(0)),
         };
         assert_eq!(h.state(), SessionState::Connecting);
     }
@@ -238,8 +254,26 @@ mod tests {
             tx: mpsc::channel(1).0,
             screen: Arc::new(Mutex::new(vt100::Parser::new(24, 80, 0))),
             state: Arc::new(AtomicU8::new(SessionState::Live as u8)),
+            revision: Arc::new(AtomicU64::new(0)),
         };
         h.end();
         assert_eq!(h.state(), SessionState::Ended);
     }
+
+    #[test]
+    fn revision_starts_at_zero_and_tracks_screen_mutations() {
+        let revision = Arc::new(AtomicU64::new(0));
+        let h = SessionHandle {
+            tx: mpsc::channel(1).0,
+            screen: Arc::new(Mutex::new(vt100::Parser::new(24, 80, 0))),
+            state: Arc::new(AtomicU8::new(SessionState::Live as u8)),
+            revision: revision.clone(),
+        };
+        assert_eq!(h.revision(), 0);
+
+        // Stands in for the WS pump: process bytes, then bump.
+        h.screen.lock().unwrap().process(b"hello");
+        revision.fetch_add(1, Ordering::Release);
+        assert_eq!(h.revision(), 1);
+    }
 }