Migrate TUI from cursive to ratatui by iho · Pull Request #3905 · mimblewimble/grin

iho · 2026-07-10T14:53:01Z

Summary

Replaces the cursive/cursive_table_view TUI (which pulls in pancurses/ncurses) with ratatui + crossterm, which have no external C dependency
Fixes the cross-compilation failures, macOS mouse-click crashes, and Windows first-run crashes described in the issue
All 5 screens (Basic Status, Peers and Sync, Mining, Logs, Version Info) and their keybindings (j/k/arrows, Tab, Enter, Esc, q, plus w/d to toggle the Mining sub-view) are preserved
New: mouse support (click a menu item to switch tabs, scroll wheel moves table selection — safe now that ncurses is gone), PageUp/PageDown/Home/End navigation in tables, and throttled redraws (repaint only on input or new data, 250ms cap) to reduce idle CPU
Interactive column sorting (cursive_table_view) is dropped, since ratatui::widgets::Table has no equivalent — tables render in the order ServerStats provides. Can be revisited as a follow-up if wanted.
Controller::new/Controller::run()'s public API (used by src/bin/cmd/server.rs) is unchanged
Windows note: key events are filtered to KeyEventKind::Press, avoiding the classic doubled-keystroke issue with crossterm on Windows

Test plan

cargo build --bin grin — 0 errors
cargo test --bin grin tui::status — both update_sync_status unit tests pass unmodified
cargo clippy --bin grin — zero findings in src/bin/tui
Confirmed via otool -L the built binary has no ncurses/pancurses linkage
Manually drove the TUI against a live testnet node in a tmux pty: all 5 tabs render with real peer/mining/sync data; keyboard navigation, mining workers/difficulty toggle, table selection clamping, Home/End/PageUp/PageDown, mouse tab-clicks and wheel scrolling (via SGR escape injection) all verified; error-dialog path renders correctly; q exits cleanly with exit code 0 and the terminal fully restored (raw mode off, alternate screen left, mouse capture disabled)

ardocrat · 2026-07-10T17:36:26Z

@@ -0,0 +1,152 @@
+// Copyright 2024 The Grin Developers


Yes, thanks

wiesche89

please rebase (merge conflicts)

Replaces the cursive/pancurses-backed TUI with ratatui + crossterm, which has no external C dependency (no more linking libncurses). This fixes cross-compilation failures, macOS crashes on mouse clicks, and Windows crashes on first run reported in mimblewimble#3841. The cursive retained-mode widget tree is replaced with a single App struct (src/bin/tui/app.rs) that every screen renders from each frame. All previous behavior (status fields, peer/mining table columns and formatting, bottom-anchored log view, keybindings, Controller's public API) is preserved, with a few additions: - Mouse support: click a menu item to switch tabs, scroll wheel moves table selection (safe now that ncurses is gone) - PageUp/PageDown/Home/End navigation in tables, with selection clamped to the table's row count - Redraws are throttled: the UI only repaints on input or new data, with a 250ms cap to keep time-based fields fresh Interactive column sorting (cursive_table_view) is dropped since ratatui's Table has no equivalent; tables render in the order ServerStats provides. Fixes mimblewimble#3841

Removes the now-unused cursive/pancurses dependency tree and resolves the ratatui/crossterm graph against the current staging lockfile.

wiesche89

Ratatui looks like the right direction and works well on macOS. I couldn’t test Linux or Windows. I left a few comments that I’d prefer to address in this PR so the migration lands cleanly.

wiesche89 · 2026-07-20T18:32:40Z

-		self.cursive.step();
+		if self.needs_redraw || self.last_draw.elapsed() >= MAX_REDRAW_INTERVAL {
+			let app = &mut self.app;
+			let _ = self.terminal.draw(|f| draw(f, app));


Could we propagate terminal input and render errors instead of dropping them? A failed poll, read, or draw currently leaves the controller running without a usable UI, and a failed draw is still marked as completed.

Done in 6f78dcb. UI::step() now returns io::Result<bool> and propagates failures from event::poll, event::read, and terminal.draw instead of swallowing them. On error the controller logs the failure, sets exit code 1, and leaves the loop so we no longer keep running with a dead UI or mark a failed draw as completed.

wiesche89 · 2026-07-20T18:34:03Z

-			if show_dialog_clone.load(Ordering::Relaxed) {
-				c.pop_layer();
+	fn handle_key(&mut self, code: KeyCode) {
+		if let Some(dialog) = &self.app.dialog {


Mouse input is blocked for every dialog, but keyboard input is only blocked for errors. During startup, Tab/j/Enter can still change the view behind the info dialog. Could we make these dialogs modal while still allowing q to shut down?

Done in 6f78dcb. All dialogs are modal now: keyboard input is blocked while either an info or error dialog is up. Error dialogs still accept Enter / q / Esc to exit; info dialogs (startup status) only allow q so you can shut down without Tab/j/Enter changing the view behind them. Mouse input was already blocked for every dialog.

wiesche89 · 2026-07-20T18:34:59Z

+	for raw_line in text.split('\n') {
+		let mut current = String::new();
+		let mut current_len = 0usize;
+		for word in raw_line.split(' ') {


Could we use Paragraph::wrap() and line_count() here instead of maintaining a separate wrapping implementation? That would preserve terminal display width and whitespace while still allowing the log view to stay bottom aligned.

Done in 6f78dcb. The custom wrap_text helper is gone. Logs now use Paragraph::wrap(Wrap { trim: false }) plus line_count() (via the unstable-rendered-line-info feature) so wrapping matches terminal display width. Bottom alignment is kept by scrolling when content overflows, or rendering into a bottom sub-area when it is shorter than the pane.

wiesche89 · 2026-07-20T18:36:16Z

+				state.select(None);
+				return;
+			}
+			let current = state.selected().unwrap_or(0) as i64;


TableState starts at None, so the first Down/j/scroll jumps straight to row 1. Could we anchor the first move to row 0 and cover it with a small test?

Done in 6f78dcb. Selection logic is in next_table_selection: when TableState is still None, the first move anchors to row 0 (or the last row for End). Unit tests cover the unselected → 0 case, empty tables, and clamp/move behavior.

wiesche89 · 2026-07-20T18:38:02Z

-         GGGGGGGGGGGGG       GGGGGGGG       GGGGGGGGGGGGG  
-        GGGGGGGGGGGGGG       GGGGGGGG       GGGGGGGGGGGGGG 
-        GGGGGGGGGGGGGG       GGGGGGGGGGGGGGGGGGGGGGGGGGGGG 
+pub const _WELCOME_LOGO: &str = "                 GGGGG                      GGGGGGG


After the migration this module only contains the unused _WELCOME_LOGO. Could we remove the file and mod constant, as part of the cleanup?

Done in 6f78dcb. Removed src/bin/tui/constants.rs and its mod constants entry; _WELCOME_LOGO was unused after the migration.

wiesche89 · 2026-07-20T18:38:58Z

 	pub fn stop(&mut self) {
-		self.cursive.quit();
+		self.app.should_quit = true;
+		let _ = disable_raw_mode();


The terminal restore sequence is now repeated in the panic hook, stop(), and Drop. Could we centralize it and make the cleanup so these paths cannot drift?

Done in 6f78dcb. Terminal teardown is centralized in restore_terminal(), shared by the panic hook, stop(), and Drop. UI also tracks a restored flag so the sequence only runs once and cannot drift between those paths.

wiesche89 · 2026-07-20T18:42:46Z

-						_ => {}
-					},
+			match message {
+				UIMessage::UpdateStatus(update) => self.app.stats = Some(update),


Now that the UI owns App and status updates happen on the same controller thread, could we update app.stats directly and remove the single-message UI channel? It looks like leftover Cursive plumbing.

Done in 6f78dcb. Removed the UIMessage channel and types.rs. Stats are written straight to app.stats on the controller thread (same thread that owns App), and the redraw flag is set there.

wiesche89 · 2026-07-20T18:54:49Z

+				lines.push(line("Stem Pool Size:", "0 (0)"));
+			}
+			lines.push(Line::from(SEPARATOR));
+			// These three lines are reserved for mining config/status/network info,


These fields were never populated in the old view. Could we drop the reserved blank rows instead of carrying the unused placeholders into the new renderer?

Done in 6f78dcb. Dropped the three reserved blank rows (and the trailing separator that only framed them) from the status view so we no longer carry unused placeholders forward.

wiesche89 · 2026-07-20T18:56:02Z

+
+/// Redraw at least this often even without input or new data, so
+/// time-based fields (peer "last seen" ages, etc.) stay fresh.
+const MAX_REDRAW_INTERVAL: Duration = Duration::from_millis(250);


The time based fields only change once per second, while 250 ms still rebuilds the full UI four times per second when idle. Could this use a one-second refresh, or is there a measurement supporting the shorter interval?

Done in 6f78dcb. Switched MAX_REDRAW_INTERVAL from 250 ms to 1 s so idle redraws match the cadence of the time-based fields (uptime, peer “last seen”, etc.) and the 1 s stats update. No measurement justified the shorter interval.

wiesche89 · 2026-07-20T18:59:29Z

+	let height = area.height as usize;

-		LogBufferView { buffer }
+	// Walk entries newest-first, wrapping each until we have enough rows to


There’s quite a bit of commentary walking through this rendering step by step. Could we simplify it by using Ratatui’s wrapping and line counting support, then keep only the non obvious bottom alignment note?

Done in 6f78dcb (together with the wrap/line_count change above). Step-by-step wrapping commentary is gone; the log draw path uses ratatui wrapping and only documents the non-obvious bottom-alignment behavior.

Propagate terminal I/O errors, make startup dialogs modal, use Paragraph wrap/line_count for logs, fix first table selection, centralize terminal restore, drop the leftover UI channel and unused constants, and align idle redraw with the 1s stats cadence.

iho changed the base branch from master to staging July 10, 2026 16:39

ardocrat reviewed Jul 10, 2026

View reviewed changes

wiesche89 reviewed Jul 19, 2026

View reviewed changes

iho added 3 commits July 20, 2026 18:28

2026

0b050a7

Regenerate Cargo.lock after rebase onto staging

0f62d93

Removes the now-unused cursive/pancurses dependency tree and resolves the ratatui/crossterm graph against the current staging lockfile.

iho force-pushed the tui-ratatui-migration branch from c08dfce to 0f62d93 Compare July 20, 2026 15:32

iho requested review from ardocrat and wiesche89 July 20, 2026 15:33

wiesche89 reviewed Jul 20, 2026

View reviewed changes

iho requested a review from wiesche89 August 1, 2026 10:31

Uh oh!

Conversation

iho commented Jul 10, 2026

Summary

Test plan

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

wiesche89 left a comment

Choose a reason for hiding this comment

Uh oh!

wiesche89 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants