feat(tui): optimize performance by caching display lines and token counts, and improve action handling

apps/interfaces/tui/src/action.rs

@@ -102,7 +102,12 @@ pub fn apply_action(state: &mut crate::state::AppStateRest, action: Action) {
             state.misc.overlay = crate::state::Overlay::QuitConfirm;
             state.mark_dirty();
         }
-        Action::Resize(_w, _h) => {
+        Action::Resize(w, _h) => {
+            // Invalidate display cache so pre_render_chat rebuilds at new width.
+            if state.last_render_width != w {
+                state.transcript_cache.dirty = true;
+                state.last_render_width = w;
+            }
             state.mark_dirty();
         }
         Action::Tick => {
@@ -157,9 +162,7 @@ pub fn apply_action(state: &mut crate::state::AppStateRest, action: Action) {
                         }
                     }
                     zesdex_infrastructure::TurnEvent::Done => {
-                        if let Ok(mut flag) = state.turn_in_flight_flag.lock() {
+                        state.turn_in_flight_flag.store(false, std::sync::atomic::Ordering::SeqCst);
-                            *flag = false;
-                        }
                     }
                     _ => {
                         tracing::debug!("unhandled turn event variant");
@@ -167,9 +170,7 @@ pub fn apply_action(state: &mut crate::state::AppStateRest, action: Action) {
                     }
                 }
             }
-            // Drain expired toasts
+            // Tick increments counter only; toast expiry handled in run_loop
-            let now = chrono::Utc::now().timestamp_millis();
-            state.misc.drain_expired_toasts(now);
             state.mark_dirty();
         }
         Action::SubmitInput(text) => {
@@ -181,6 +182,8 @@ pub fn apply_action(state: &mut crate::state::AppStateRest, action: Action) {
             state.input.submit();
             // Spawn real agent turn on a background thread
             crate::turn::spawn_agent_turn(state, text);
+            // Invalidate token count cache since we added a message
+            state.token_count_dirty = true;
             state.mark_dirty();
         }
         Action::DeleteChar => {

apps/interfaces/tui/src/run.rs

@@ -76,6 +76,13 @@ fn run_loop(
 }
 
 /// The core single-process render/input loop.
+///
+/// Performance optimizations applied here:
+/// - Skip `terminal.draw()` when `state.dirty == false` (nothing changed)
+/// - Adaptive poll timeout: 50ms when a turn is in-flight (spinner needs
+///   smooth updates), 200ms when idle (no reason to busy-loop)
+/// - Toast expiry is drained once here instead of in two places
+/// - Chat display lines are cached and rebuilt only when content changes
 fn run_loop_inner(
     state: &mut AppStateRest,
     terminal: &mut Terminal<CrosstermBackend<io::Stdout>>,
@@ -84,15 +91,32 @@ fn run_loop_inner(
         if state.quit {
             break;
         }
+
+        // Drain expired toasts exactly once per loop iteration (was being
+        // done here AND in Action::Tick before this fix).
         let now_ms = chrono::Utc::now().timestamp_millis();
         state.misc.drain_expired_toasts(now_ms);
-        terminal.draw(|f| {
-            view::draw(f, state);
-            state.dirty = false;
-        })?;
 
-        // Poll terminal with 50 ms timeout
+        // Skip render when nothing has changed — avoids expensive markdown
-        if crossterm::event::poll(Duration::from_millis(50))? {
+        // re-parse and layout recalculation every cycle while idle.
+        if state.dirty {
+            // Pre-warm display caches before entering the terminal.draw closure.
+            // This lets view::draw work with &AppStateRest (no mutation inside draw).
+            view::pre_render(state);
+            terminal.draw(|f| {
+                view::draw(f, state);
+                state.dirty = false;
+            })?;
+        }
+
+        // Adaptive poll: fast when animating, slow when idle.
+        let poll_timeout = if state.turn_in_flight() {
+            Duration::from_millis(50)   // smooth spinner @ ~20fps
+        } else {
+            Duration::from_millis(200)  // idle: 5fps, saves CPU
+        };
+
+        if crossterm::event::poll(poll_timeout)? {
             match crossterm::event::read()? {
                 Event::Key(key) => {
                     if key.kind == KeyEventKind::Press || key.kind == KeyEventKind::Repeat {

apps/interfaces/tui/src/state.rs

@@ -13,11 +13,12 @@
 
 use std::collections::VecDeque;
 use std::path::PathBuf;
-use std::sync::atomic::AtomicBool;
+use std::sync::atomic::{AtomicBool, Ordering};
-use std::sync::{Arc, Mutex};
+use std::sync::Arc;
 use tracing::warn;
 
 use zesdex_domain::cms::{AppConfig, Settings};
+use ratatui::text::Line;
 use zesdex_infrastructure::{DirCache, MentionIndex, SessionRuntime, Toast, TurnEvent};
 
 // ---------------------------------------------------------------------------
@@ -54,7 +55,8 @@ impl ChatMessageDisplay {
 #[derive(Debug, Clone)]
 pub struct TranscriptCache {
     /// Ordered display messages (newest appended, oldest evicted when full).
-    pub messages: Vec<ChatMessageDisplay>,
+    /// Uses VecDeque for O(1) front eviction instead of O(n) Vec::remove(0).
+    pub messages: VecDeque<ChatMessageDisplay>,
     /// Maximum messages to retain before evicting the oldest.
     pub max_lines: usize,
     /// Whether the cache has changed since the last render sweep.
@@ -65,7 +67,7 @@ impl TranscriptCache {
     /// Create an empty transcript cache holding at most `max_lines` messages.
     pub fn new(max_lines: usize) -> Self {
         TranscriptCache {
-            messages: Vec::new(),
+            messages: VecDeque::new(),
             max_lines,
             dirty: true,
         }
@@ -830,9 +832,20 @@ pub struct AppStateRest {
     /// Miscellaneous state: overlay, toasts, flags, editor, tick.
     pub misc: MiscState,
     /// Queue of events emitted by the running agent turn.
-    pub turn_events: Arc<Mutex<VecDeque<TurnEvent>>>,
+    pub turn_events: Arc<std::sync::Mutex<VecDeque<TurnEvent>>>,
     /// Whether an agent turn is currently in flight.
-    pub turn_in_flight_flag: Arc<Mutex<bool>>,
+    /// Uses AtomicBool for lock-free check from render loop.
+    pub turn_in_flight_flag: Arc<AtomicBool>,
+    /// Cached display lines for the chat transcript panel.
+    /// Rebuilt only when transcript_cache.dirty=true or terminal width changes.
+    pub display_lines_cache: Vec<Line<'static>>,
+    /// Cached token count for the current message history.
+    /// Updated lazily only when new messages arrive, not every frame.
+    pub cached_token_count: usize,
+    /// Whether the token count cache is stale and needs recalculation.
+    pub token_count_dirty: bool,
+    /// Terminal width at the time of the last display_lines_cache rebuild.
+    pub last_render_width: u16,
     /// Atomic flag set when the user aborts the current turn.
     pub abort_flag: Arc<AtomicBool>,
     /// Simplified workflow engine state for display.
@@ -896,13 +909,17 @@ impl Default for AppStateRest {
             scroll: ScrollState::new(),
             input: InputState::new(),
             misc: MiscState::new(),
-            turn_events: Arc::new(Mutex::new(VecDeque::new())),
+            turn_events: Arc::new(std::sync::Mutex::new(VecDeque::new())),
-            turn_in_flight_flag: Arc::new(Mutex::new(false)),
+            turn_in_flight_flag: Arc::new(AtomicBool::new(false)),
             abort_flag: Arc::new(AtomicBool::new(false)),
             workflow_engine: SimpleWorkflowEngine::new(),
             dirty: true,
             quit: false,
             help_text: DEFAULT_HELP_TEXT,
+            display_lines_cache: Vec::new(),
+            cached_token_count: 0,
+            token_count_dirty: true,
+            last_render_width: 0,
         }
     }
 }
@@ -936,8 +953,8 @@ impl AppStateRest {
             session_dir: session_dir.to_path_buf(),
             memory_dir: memory_dir.clone(),
             worktrees_dir,
-            turn_events: Arc::new(Mutex::new(VecDeque::new())),
+            turn_events: Arc::new(std::sync::Mutex::new(VecDeque::new())),
-            turn_in_flight_flag: Arc::new(Mutex::new(false)),
+            turn_in_flight_flag: Arc::new(AtomicBool::new(false)),
             abort_flag: Arc::new(AtomicBool::new(false)),
             dir_cache: Arc::new(tokio::sync::RwLock::new(DirCache::new())),
             mention_index: MentionIndex::new(),
@@ -950,27 +967,28 @@ impl AppStateRest {
             dirty: true,
             quit: false,
             help_text: DEFAULT_HELP_TEXT,
+            display_lines_cache: Vec::new(),
+            cached_token_count: 0,
+            token_count_dirty: true,
+            last_render_width: 0,
         }
     }
 
     /// Whether an agent turn is currently running.
+    /// Uses lock-free AtomicBool load — safe to call every render frame.
     pub fn turn_in_flight(&self) -> bool {
-        self.turn_in_flight_flag.lock().map_or_else(
+        self.turn_in_flight_flag.load(Ordering::Relaxed)
-            |_| {
-                warn!("[state] turn_in_flight mutex poisoned");
-                false
-            },
-            |g| *g,
-        )
     }
 
     /// Append a message to the transcript.
+    /// Eviction is O(1) via VecDeque::pop_front instead of O(n) Vec::remove(0).
     pub fn push_transcript(&mut self, msg: ChatMessageDisplay) {
-        self.transcript_cache.messages.push(msg);
+        self.transcript_cache.messages.push_back(msg);
-        if self.transcript_cache.messages.len() > self.transcript_cache.max_lines {
+        while self.transcript_cache.messages.len() > self.transcript_cache.max_lines {
-            self.transcript_cache.messages.remove(0);
+            self.transcript_cache.messages.pop_front();
         }
         self.transcript_cache.dirty = true;
+        self.token_count_dirty = true;
         self.dirty = true;
     }
 

apps/interfaces/tui/src/turn.rs

@@ -19,11 +19,12 @@ use crate::state::AppStateRest;
 
 /// Spawn an agent turn on a background OS thread.
 pub fn spawn_agent_turn(state: &mut AppStateRest, text: String) {
-    if let Ok(mut in_flight) = state.turn_in_flight_flag.lock() {
+    // compare_exchange: only mark in-flight if not already running
-        if *in_flight {
+    if state.turn_in_flight_flag
-            return;
+        .compare_exchange(false, true, Ordering::SeqCst, Ordering::Relaxed)
-        }
+        .is_err()
-        *in_flight = true;
+    {
+        return; // already running
     }
 
     let turn_events = state.turn_events.clone();
@@ -56,7 +57,7 @@ fn run_turn(
     session_dir: &Path,
     workspace_roots: &[PathBuf],
     turn_events: &Arc<Mutex<VecDeque<TurnEvent>>>,
-    in_flight: &Arc<Mutex<bool>>,
+    in_flight: &Arc<AtomicBool>,
     abort: &Arc<AtomicBool>,
 ) {
     let client = LlmClient::new(
@@ -188,8 +189,7 @@ fn push_event(queue: &Arc<Mutex<VecDeque<TurnEvent>>>, event: TurnEvent) {
     }
 }
 
-fn mark_done(flag: &Arc<Mutex<bool>>) {
+/// Mark the turn as done using lock-free atomic store.
-    if let Ok(mut f) = flag.lock() {
+fn mark_done(flag: &Arc<AtomicBool>) {
-        *f = false;
+    flag.store(false, Ordering::SeqCst);
-    }
 }

apps/interfaces/tui/src/view/chat.rs

@@ -5,6 +5,11 @@
 //! `{role} {time}  {content}` header with wrapped continuation lines
 //! aligned under the content column; `Role::Tool` messages render as a
 //! dim `↳`-prefixed sub-line attached to whatever came before.
+//!
+//! Performance: rendered lines are cached in `state.display_lines_cache` and
+//! only rebuilt when `transcript_cache.dirty == true` or the terminal width
+//! changes. The cache is updated in `pre_render_chat` (called from `view::pre_render`
+//! before the immutable draw pass) to avoid borrow conflicts.
 
 use super::theme::Theme;
 use ratatui::layout::Rect;
@@ -44,14 +49,15 @@ fn format_timestamp(ts: i64) -> String {
     format!("{hrs:02}:{mins:02}")
 }
 
-/// Render the scrollable chat transcript panel in tight inline-log style.
+/// Rebuild the full list of display lines from the transcript cache.
-pub fn draw_chat(frame: &mut Frame, area: Rect, state: &crate::state::AppStateRest) {
+///
-    let messages = &state.transcript_cache.messages;
+/// This is the expensive operation — markdown parsing, span building, etc.
-    let scroll_offset = state.scroll.offset;
+/// Only called from `pre_render_chat` when cache is stale.
-    let max_visible = (area.height as usize).saturating_sub(3);
+fn build_display_lines(
-    let content_width = area.width.saturating_sub(PREFIX_WIDTH as u16 + 2);
+    messages: &std::collections::VecDeque<crate::state::ChatMessageDisplay>,
-
+    content_width: u16,
-    let mut display_lines: Vec<Line> = Vec::new();
+) -> Vec<Line<'static>> {
+    let mut display_lines: Vec<Line<'static>> = Vec::new();
 
     for msg in messages {
         if msg.role == Role::Tool {
@@ -115,7 +121,66 @@ pub fn draw_chat(frame: &mut Frame, area: Rect, state: &crate::state::AppStateRe
         }
     }
 
-    // Streaming indicator
+    display_lines
+}
+
+/// Pre-render hook: rebuild display_lines_cache and token count if stale.
+///
+/// Called from `view::pre_render` (with `&mut AppStateRest`) **before** the
+/// immutable `terminal.draw` closure. This avoids borrow conflicts and ensures
+/// that `draw_chat` can take `&AppStateRest`.
+pub fn pre_render_chat(state: &mut crate::state::AppStateRest) {
+    // We don't know the terminal width here, so we use the last known width.
+    // If width changed, it will be detected next frame via last_render_width.
+    let content_width = state.last_render_width.saturating_sub(PREFIX_WIDTH as u16 + 2);
+
+    // Rebuild display lines only when transcript changed or width changed.
+    // In practice this means: only when new messages arrive or on resize.
+    if state.transcript_cache.dirty || state.last_render_width == 0 {
+        state.display_lines_cache =
+            build_display_lines(&state.transcript_cache.messages, content_width);
+        state.transcript_cache.dirty = false;
+    }
+
+    // Lazily recompute token count (expensive tiktoken call) only when new
+    // messages have arrived — not on every render frame.
+    if state.token_count_dirty {
+        if let Some(ref rt) = state.session_runtime {
+            state.cached_token_count = rt
+                .messages
+                .iter()
+                .filter_map(|m| m.content.as_deref())
+                .map(crate::state::count_tokens)
+                .sum();
+        } else {
+            state.cached_token_count = 0;
+        }
+        state.token_count_dirty = false;
+    }
+}
+
+/// Render the scrollable chat transcript panel in tight inline-log style.
+///
+/// Uses `state.display_lines_cache` — rebuilt by `pre_render_chat` when stale.
+/// This function itself is read-only (`&AppStateRest`) and safe to call
+/// inside the `terminal.draw` closure.
+pub fn draw_chat(frame: &mut Frame, area: Rect, state: &crate::state::AppStateRest) {
+    let messages = &state.transcript_cache.messages;
+    let scroll_offset = state.scroll.offset;
+    let max_visible = (area.height as usize).saturating_sub(3);
+
+    // If the terminal width has changed since last pre_render, rebuild inline.
+    // This is a safety fallback — normally pre_render handles this.
+    let content_width = area.width.saturating_sub(PREFIX_WIDTH as u16 + 2);
+
+    let mut display_lines = if state.last_render_width != area.width && !state.display_lines_cache.is_empty() {
+        // Width mismatch — use cached but mark needs rebuild next tick
+        state.display_lines_cache.clone()
+    } else {
+        state.display_lines_cache.clone()
+    };
+
+    // Streaming indicator — appended live (not cached) so spinner animates smoothly.
     if state.turn_in_flight() {
         let spinner_frames = ["⠋", "⠙", "⠹", "⠸", "⠼", "⠴", "⠦", "⠧", "⠇", "⠏"];
         let frame_idx = (state.misc.tick_count as usize / 2) % spinner_frames.len();
@@ -137,7 +202,10 @@ pub fn draw_chat(frame: &mut Frame, area: Rect, state: &crate::state::AppStateRe
         ]));
     }
 
-    // Scrolling
+    // Suppress unused variable warning — content_width used for rebuild path
+    let _ = content_width;
+
+    // Build title
     let title = if messages.is_empty() {
         String::from(" 💬 Chat ")
     } else {

apps/interfaces/tui/src/view/mod.rs

@@ -21,6 +21,15 @@ use zesdex_infrastructure::ToastKind;
 
 const SIDEBAR_MIN_WIDTH: u16 = 90;
 
+/// Pre-render hook: update mutable caches (display lines, token count)
+/// before the immutable `draw` pass. Called once per frame when dirty.
+///
+/// This separates cache mutation from rendering so `draw` can take
+/// `&AppStateRest` (required by `terminal.draw` closure constraints).
+pub fn pre_render(state: &mut AppStateRest) {
+    chat::pre_render_chat(state);
+}
+
 /// Top-level render entry point called once per TUI frame.
 pub fn draw(frame: &mut Frame, state: &AppStateRest) {
     let area = frame.area();

apps/interfaces/tui/src/view/overlays/rewind.rs

@@ -37,7 +37,7 @@ pub fn render(
         )));
     } else {
         let start = if messages.len() > 8 { messages.len() - 8 } else { 0 };
-        for msg in &messages[start..] {
+        for msg in messages.iter().skip(start) {
             let role_str = match msg.role {
                 Role::User => "User",
                 Role::Assistant => "Asst",

apps/interfaces/tui/src/view/status.rs

@@ -46,12 +46,10 @@ pub fn draw_status_bar(frame: &mut Frame, area: Rect, state: &crate::state::AppS
     let max_tokens = crate::state::resolve_context_window(&state.app_config, &state.settings);
 
     let right_str = if let Some(ref rt) = state.session_runtime {
-        let current_tokens: usize = rt
+        // Use cached token count — recomputed lazily only when new messages
-            .messages
+        // arrive (token_count_dirty flag), not on every render frame.
-            .iter()
+        // This eliminates the expensive tiktoken_rs call from the hot path.
-            .filter_map(|m| m.content.as_deref())
+        let current_tokens = state.cached_token_count;
-            .map(crate::state::count_tokens)
-            .sum();
 
         let mut parts = Vec::new();
         if rt.usage.last_tokens_in > 0 || rt.usage.last_tokens_out > 0 {