@oh-my-pi/pi-tui 15.10.3 → 15.10.4

package/CHANGELOG.md

@@ -2,6 +2,12 @@
 
 ## [Unreleased]
 
+## [15.10.4] - 2026-06-08
+
+### Fixed
+
+- Fixed Windows ConPTY session-resume painting the transcript with the last several rows truncated below the viewport until Alt+Tab forced a host repaint. After `sessionReplace`/`historyRebuild`/`overlayRebuild` paints that scroll-push content into native scrollback, the renderer now arms a 150 ms ConPTY settle window that coalesces spinner/blink-driven `requestRender(false)` calls into a single trailing render — Windows Terminal's viewport-follow logic no longer falls further behind the cursor on every tick of the post-paint storm. The arm also reclaims any render request queued *during* the in-flight composition (notably `ImageBudget.endPass()` calling `requestRender()` synchronously when a frame trips the live-graphics cap): without that, the queued request sat on the standard 30 Hz throttle and fired at ~33 ms — well inside the 150 ms quiet window — defeating the coalescing. Bumped the ConPTY per-`WriteFile` chunk cap from 8 KiB to 16 KiB so a multi-megabyte resume paint emits half as many writes (still well under the ~32 KiB threshold from #2034 that the original cap defends against), and made the cap measure encoded UTF-8 bytes instead of JS code units so a CJK-heavy transcript can't silently inflate a 16-KiB-of-code-units chunk into ~48 KiB of `WriteFile` traffic and reintroduce the #2034 viewport bug ([#2095](https://github.com/can1357/oh-my-pi/issues/2095)).
+
 ## [15.10.3] - 2026-06-08
 
 ### Fixed

package/dist/types/terminal.d.ts

@@ -1,17 +1,24 @@
 /**
- * Split `data` into chunks no larger than `maxChunkSize`, preferring a line
- * boundary (`\n`) as the cut point so escape sequences (which never contain
- * `\n`) stay intact. The TUI's full-paint buffers are line-structured
- * (`buffer += "\r\n"` between rows), so a newline almost always exists within
- * the window. The fallback for a buffer with no newline in range is a hard
- * cut at `maxChunkSize`: the ConPTY viewport bug from a single oversized
- * write is strictly worse than a one-frame escape-sequence glitch on a buffer
- * the renderer effectively never produces.
+ * Split `data` into chunks whose encoded UTF-8 byte length is no greater than
+ * `maxChunkBytes`, preferring a line boundary (`\n`) as the cut point so
+ * escape sequences (which never contain `\n`) stay intact. The TUI's
+ * full-paint buffers are line-structured (`buffer += "\r\n"` between rows),
+ * so a newline almost always exists within the window. The fallback for a
+ * buffer with no newline in range is a hard cut at the last UTF-8 code-point
+ * boundary that still fits — the ConPTY viewport bug from a single oversized
+ * write is strictly worse than a one-frame escape-sequence glitch on a
+ * buffer the renderer effectively never produces.
+ *
+ * UTF-16 code units are walked manually rather than measuring with
+ * `Buffer.byteLength` per slice candidate: each code unit's UTF-8 width is
+ * known from its value (BMP `<0x80` → 1, `<0x800` → 2, surrogate pair → 4
+ * bytes across two units, other BMP → 3), and surrogate pairs are kept
+ * together so the chunker never splits a non-BMP character.
  *
  * Exported for unit testing of the chunking contract; `#safeWrite` is the
  * sole production caller.
  */
-export declare function chunkForConPTY(data: string, maxChunkSize?: number): string[];
+export declare function chunkForConPTY(data: string, maxChunkBytes?: number): string[];
 /**
  * Emergency terminal restore - call this from signal/crash handlers
  * Resets terminal state without requiring access to the ProcessTerminal instance
@@ -91,6 +98,15 @@ export interface Terminal {
      */
     onPrivateModeReport?(callback: (mode: number, supported: boolean) => void): void;
 }
+/**
+ * True when stdout flows through a ConPTY pseudo-console (native win32, or
+ * Linux running under WSL where stdout still crosses into ConPTY at the
+ * `wslhost` boundary). ConPTY hosts share the per-WriteFile viewport-tracking
+ * quirks documented above and on {@link MAX_CONPTY_WRITE_CHUNK_BYTES}, so both
+ * `#safeWrite` and the renderer's post-big-paint settle gate hang off this
+ * single predicate.
+ */
+export declare function isConPTYHosted(): boolean;
 /**
  * Real terminal using process.stdin/stdout
  */

package/dist/types/tui.d.ts

@@ -1,5 +1,5 @@
 import { ImageBudget } from "./components/image";
-import type { Terminal } from "./terminal";
+import { type Terminal } from "./terminal";
 import { visibleWidth } from "./utils";
 type InputListenerResult = {
     consume?: boolean;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-tui",
-	"version": "15.10.3",
+	"version": "15.10.4",
 	"description": "Terminal User Interface library with differential rendering for efficient text-based applications",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -37,8 +37,8 @@
 		"fmt": "biome format --write ."
 	},
 	"dependencies": {
-		"@oh-my-pi/pi-natives": "15.10.3",
-		"@oh-my-pi/pi-utils": "15.10.3",
+		"@oh-my-pi/pi-natives": "15.10.4",
+		"@oh-my-pi/pi-utils": "15.10.4",
 		"lru-cache": "11.5.1",
 		"marked": "^18.0.4"
 	},

package/src/terminal.ts

@@ -10,7 +10,7 @@ const TERMINAL_PROGRESS_ACTIVE_SEQUENCE = "\x1b]9;4;3\x07";
 const TERMINAL_PROGRESS_CLEAR_SEQUENCE = "\x1b]9;4;0;\x07";
 
 /**
- * Maximum bytes per `process.stdout.write` call on Windows.
+ * Maximum encoded UTF-8 bytes per `process.stdout.write` call on Windows.
  *
  * Windows ConPTY ties viewport tracking to per-`WriteFile` boundaries: when a
  * single write exceeds ~32-64 KB, the pseudo-console stops following the
@@ -20,43 +20,96 @@ const TERMINAL_PROGRESS_CLEAR_SEQUENCE = "\x1b]9;4;0;\x07";
  * first ~30 lines until any focus event forces the host to re-query the
  * cursor. The data is delivered correctly — it's purely a viewport-sync bug.
  *
- * 8 KiB is well below the 32 KiB threshold reported on Windows Terminal and
- * leaves headroom for the other ConPTY hosts (Tabby, Hyper, VS Code) where
- * the exact limit is undocumented. The cost is a handful of extra syscalls
- * per full paint — invisible compared to the cost of the paint itself.
+ * The cap is on **encoded UTF-8 bytes**, not JS code units, because
+ * `process.stdout.write(string)` UTF-8-encodes before handing off to
+ * `WriteFile`. A pure-CJK transcript row encodes to ~3 bytes per BMP code
+ * unit, so a code-unit-based cap of 16 KiB could land at ~48 KiB of actual
+ * `WriteFile` traffic and reintroduce the #2034 parked-viewport bug for
+ * non-ASCII content.
+ *
+ * 16 KiB is half the smallest observed Windows Terminal threshold (32 KiB),
+ * which keeps the per-write parked-viewport bug fixed by #2034 while halving
+ * the WriteFile count on multi-megabyte paints (a 3 MB session resume splits
+ * into ~192 chunks instead of ~384). Fewer WriteFiles means fewer chances for
+ * WT's viewport-following logic to lose track of the cursor during the burst,
+ * which mitigates the residual mid-paint drift the original 8 KiB cap left
+ * behind (#2095). Still well clear of the threshold so the other ConPTY hosts
+ * (Tabby, Hyper, VS Code) — where the exact limit is undocumented — keep
+ * their safety margin.
  */
-const MAX_CONPTY_WRITE_CHUNK = 8 * 1024;
+const MAX_CONPTY_WRITE_CHUNK_BYTES = 16 * 1024;
 
 /**
- * Split `data` into chunks no larger than `maxChunkSize`, preferring a line
- * boundary (`\n`) as the cut point so escape sequences (which never contain
- * `\n`) stay intact. The TUI's full-paint buffers are line-structured
- * (`buffer += "\r\n"` between rows), so a newline almost always exists within
- * the window. The fallback for a buffer with no newline in range is a hard
- * cut at `maxChunkSize`: the ConPTY viewport bug from a single oversized
- * write is strictly worse than a one-frame escape-sequence glitch on a buffer
- * the renderer effectively never produces.
+ * Split `data` into chunks whose encoded UTF-8 byte length is no greater than
+ * `maxChunkBytes`, preferring a line boundary (`\n`) as the cut point so
+ * escape sequences (which never contain `\n`) stay intact. The TUI's
+ * full-paint buffers are line-structured (`buffer += "\r\n"` between rows),
+ * so a newline almost always exists within the window. The fallback for a
+ * buffer with no newline in range is a hard cut at the last UTF-8 code-point
+ * boundary that still fits — the ConPTY viewport bug from a single oversized
+ * write is strictly worse than a one-frame escape-sequence glitch on a
+ * buffer the renderer effectively never produces.
+ *
+ * UTF-16 code units are walked manually rather than measuring with
+ * `Buffer.byteLength` per slice candidate: each code unit's UTF-8 width is
+ * known from its value (BMP `<0x80` → 1, `<0x800` → 2, surrogate pair → 4
+ * bytes across two units, other BMP → 3), and surrogate pairs are kept
+ * together so the chunker never splits a non-BMP character.
  *
  * Exported for unit testing of the chunking contract; `#safeWrite` is the
  * sole production caller.
  */
-export function chunkForConPTY(data: string, maxChunkSize: number = MAX_CONPTY_WRITE_CHUNK): string[] {
-	if (data.length <= maxChunkSize) return [data];
+export function chunkForConPTY(data: string, maxChunkBytes: number = MAX_CONPTY_WRITE_CHUNK_BYTES): string[] {
+	// Fast path: whole buffer fits in one write.
+	if (Buffer.byteLength(data, "utf8") <= maxChunkBytes) return [data];
 	const chunks: string[] = [];
+	const len = data.length;
 	let pos = 0;
-	while (pos < data.length) {
-		const remaining = data.length - pos;
-		if (remaining <= maxChunkSize) {
+	while (pos < len) {
+		let bytes = 0;
+		// Index just past the most recent `\n` we've consumed inside [pos, i):
+		// the natural cut point that leaves escape sequences intact.
+		let lastNewlineEnd = -1;
+		let i = pos;
+		while (i < len) {
+			const cu = data.charCodeAt(i);
+			let cuLen = 1;
+			let cuBytes: number;
+			if (cu < 0x80) {
+				cuBytes = 1;
+			} else if (cu < 0x800) {
+				cuBytes = 2;
+			} else if (cu >= 0xd800 && cu < 0xdc00) {
+				// High surrogate: pair with the following low surrogate (4 bytes
+				// across two code units); an unpaired surrogate UTF-8-encodes as
+				// the 3-byte U+FFFD replacement character.
+				const next = i + 1 < len ? data.charCodeAt(i + 1) : 0;
+				if (next >= 0xdc00 && next < 0xe000) {
+					cuBytes = 4;
+					cuLen = 2;
+				} else {
+					cuBytes = 3;
+				}
+			} else {
+				// BMP non-surrogate or unpaired low surrogate → 3 bytes.
+				cuBytes = 3;
+			}
+			if (bytes + cuBytes > maxChunkBytes && i > pos) {
+				// Would overflow the cap. Cut at the last newline if we found one,
+				// otherwise hard-cut at the current code-point boundary.
+				const cut = lastNewlineEnd > pos ? lastNewlineEnd : i;
+				chunks.push(data.slice(pos, cut));
+				pos = cut;
+				break;
+			}
+			bytes += cuBytes;
+			i += cuLen;
+			if (cu === 0x0a) lastNewlineEnd = i;
+		}
+		if (i >= len) {
 			chunks.push(data.slice(pos));
-			break;
+			pos = len;
 		}
-		const windowEnd = pos + maxChunkSize;
-		// Prefer the last newline inside the window so escape sequences stay
-		// intact within their chunk; hard-cut at `windowEnd` otherwise.
-		const nl = data.lastIndexOf("\n", windowEnd - 1);
-		const cut = nl >= pos ? nl + 1 : windowEnd;
-		chunks.push(data.slice(pos, cut));
-		pos = cut;
 	}
 	return chunks;
 }
@@ -202,7 +255,17 @@ export interface Terminal {
 	onPrivateModeReport?(callback: (mode: number, supported: boolean) => void): void;
 }
 
-function isWindowsSubsystemForLinux(): boolean {
+/**
+ * True when stdout flows through a ConPTY pseudo-console (native win32, or
+ * Linux running under WSL where stdout still crosses into ConPTY at the
+ * `wslhost` boundary). ConPTY hosts share the per-WriteFile viewport-tracking
+ * quirks documented above and on {@link MAX_CONPTY_WRITE_CHUNK_BYTES}, so both
+ * `#safeWrite` and the renderer's post-big-paint settle gate hang off this
+ * single predicate.
+ */
+export function isConPTYHosted(): boolean {
+	if (process.platform === "win32") return true;
+	// WSL: stdout still crosses into ConPTY at the `wslhost` boundary.
 	return process.platform === "linux" && (!!$env.WSL_DISTRO_NAME || !!$env.WSL_INTEROP);
 }
 
@@ -349,7 +412,8 @@ export class ProcessTerminal implements Terminal {
 		// Windows Terminal under WSL has been observed to close the hosting tab
 		// after repeated OSC 11/DA1 probes. Keep the initial/event-driven probes,
 		// but avoid background polling there.
-		if (!isWindowsSubsystemForLinux()) {
+		const isWSL = process.platform === "linux" && (!!$env.WSL_DISTRO_NAME || !!$env.WSL_INTEROP);
+		if (!isWSL) {
 			this.#startOsc11Poll();
 		}
 
@@ -1090,10 +1154,12 @@ export class ProcessTerminal implements Terminal {
 			// WSL — `process.platform === "linux"` there, but stdout still
 			// crosses into ConPTY at the `wslhost` boundary, so the same per-
 			// WriteFile cap applies. Non-ConPTY PTYs keep the single-write fast
-			// path. See #2034.
-			const conptyHosted = process.platform === "win32" || isWindowsSubsystemForLinux();
-			if (conptyHosted && data.length > MAX_CONPTY_WRITE_CHUNK) {
-				for (const chunk of chunkForConPTY(data, MAX_CONPTY_WRITE_CHUNK)) {
+			// path. The cap is on encoded UTF-8 bytes, not JS code units, because
+			// `process.stdout.write(string)` UTF-8-encodes before `WriteFile`,
+			// and a code-unit cap would let CJK transcript rows expand past the
+			// threshold. See #2034 and #2095.
+			if (isConPTYHosted() && Buffer.byteLength(data, "utf8") > MAX_CONPTY_WRITE_CHUNK_BYTES) {
+				for (const chunk of chunkForConPTY(data, MAX_CONPTY_WRITE_CHUNK_BYTES)) {
 					process.stdout.write(chunk);
 				}
 			} else {

package/src/tui.ts

@@ -17,7 +17,7 @@ import { $flag, getDebugLogPath } from "@oh-my-pi/pi-utils";
 import { DEFAULT_MAX_INLINE_IMAGES, ImageBudget } from "./components/image";
 import { planDeccaraFills } from "./deccara";
 import { isKeyRelease, matchesKey } from "./keys";
-import type { Terminal } from "./terminal";
+import { isConPTYHosted, type Terminal } from "./terminal";
 import {
 	encodeKittyDeleteImage,
 	ImageProtocol,
@@ -494,6 +494,26 @@ export class TUI extends Container {
 	// arrives (issue #2088). Coalescing every SIGWINCH inside this window into
 	// a single forced render lets the multiplexer settle first.
 	static readonly #MULTIPLEXER_RESIZE_DEBOUNCE_MS = 50;
+	// Post-paint settle window for ConPTY hosts. The `sessionReplace` /
+	// `historyRebuild` / `overlayRebuild` intents drive `#emitFullPaint` over
+	// a transcript that overflows the viewport, scroll-pushing everything past
+	// the last `height` rows into native scrollback. Windows Terminal's
+	// viewport-follow logic gets lossy during that burst: spinner/blink-driven
+	// `requestRender(false)` calls firing inside the window each produce another
+	// diff write, and the WT host processes them faster than its viewport
+	// tracker can keep up — the visible tail ends up parked a few rows above
+	// the actual last row until any focus event (Alt+Tab) forces a host repaint.
+	// Coalescing every non-forced render inside this window into a single
+	// trailing render lets the host fully settle the big paint before any
+	// follow-up writes touch the buffer. The first-ever `initial` paint is
+	// deliberately exempt: nothing has been on screen yet, so no drift can
+	// have accumulated, and tests that start the TUI over an over-tall
+	// component depend on the next paint firing without delay. Only armed on
+	// ConPTY hosts (`isConPTYHosted()`); other terminals do not exhibit the
+	// drift and would just see an unnecessary post-paint latency. See #2095.
+	static readonly #CONPTY_POST_FULL_PAINT_SETTLE_MS = 150;
+	#postFullPaintSettleUntilMs = 0;
+	#postFullPaintSettleTimer: RenderTimer | undefined;
 	#cursorRow = 0; // Logical cursor row (end of rendered content)
 	#hardwareCursorRow = 0; // Actual terminal cursor row (may differ due to IME positioning)
 	#hardwareCursorState: HardwareCursorState | null = null;
@@ -1106,6 +1126,7 @@ export class TUI extends Container {
 			this.#multiplexerResizeTimer.cancel();
 			this.#multiplexerResizeTimer = undefined;
 		}
+		this.#clearPostFullPaintSettle();
 		this.#deferredForcedClearScrollback = false;
 		// Place the parent shell on the first line after the rendered content. When
 		// that line is still inside the viewport, moving there and writing `\r` is
@@ -1214,6 +1235,10 @@ export class TUI extends Container {
 				this.#armMultiplexerResizeTimer(options?.clearScrollback === true);
 				return;
 			}
+			// A forced render preempts the post-full-paint ConPTY settle: it owns
+			// the next paint and is going to redraw the buffer anyway, so the
+			// trailing coalesced render queued by the settle would only race it.
+			this.#clearPostFullPaintSettle();
 			this.#prepareForcedRender(options?.clearScrollback === true);
 			this.#renderRequested = true;
 			this.#renderScheduler.scheduleImmediate(() => {
@@ -1226,6 +1251,27 @@ export class TUI extends Container {
 			});
 			return;
 		}
+		// Coalesce non-forced renders inside the post-full-paint ConPTY settle
+		// window into one trailing render. Spinner/blink/streaming components
+		// otherwise fire `requestRender(false)` at 30 Hz while the host is still
+		// catching up with the previous big paint, and each follow-up viewport
+		// repaint nudges Windows Terminal's viewport tracker further off the
+		// last row (see #2095).
+		if (this.#postFullPaintSettleUntilMs > 0) {
+			const now = this.#renderScheduler.now();
+			if (now < this.#postFullPaintSettleUntilMs) {
+				if (this.#postFullPaintSettleTimer === undefined) {
+					this.#postFullPaintSettleTimer = this.#renderScheduler.scheduleRender(() => {
+						this.#postFullPaintSettleTimer = undefined;
+						this.#postFullPaintSettleUntilMs = 0;
+						if (this.#stopped) return;
+						this.requestRender(false);
+					}, this.#postFullPaintSettleUntilMs - now);
+				}
+				return;
+			}
+			this.#postFullPaintSettleUntilMs = 0;
+		}
 		if (this.#renderRequested) return;
 		this.#renderRequested = true;
 		this.#renderScheduler.scheduleImmediate(() => this.#scheduleRender());
@@ -1262,6 +1308,64 @@ export class TUI extends Container {
 			this.requestRender(true, { clearScrollback: deferredClearScrollback });
 		}, TUI.#MULTIPLEXER_RESIZE_DEBOUNCE_MS);
 	}
+
+	/**
+	 * Arm the post-full-paint settle window after an `#emitFullPaint` that
+	 * pushed content into native scrollback on a ConPTY host. Idempotent inside
+	 * the window: a later overflowing paint extends `until` to the later
+	 * deadline so back-to-back big paints do not double-fire the trailing
+	 * coalesced render, and the existing deferred timer is rescheduled to the
+	 * later deadline.
+	 *
+	 * Mid-composition callers (most notably `ImageBudget.endPass()`, which can
+	 * call `requestRender()` from inside the in-flight paint when a new image
+	 * trips the budget) queue their render *before* the settle exists, so they
+	 * fall through the gate and set `#renderRequested` / `#renderTimer` on the
+	 * 30 Hz throttle. Without absorbing those, the throttled follow-up fires
+	 * inside the 150 ms quiet window and reintroduces the cascade the settle
+	 * was meant to stop. Cancel both, then eagerly arm the trailing settle
+	 * timer so the in-flight request still rides one coalesced render at the
+	 * end of the window. See #2095.
+	 */
+	#armPostFullPaintSettle(): void {
+		if (!isConPTYHosted()) return;
+		const until = this.#renderScheduler.now() + TUI.#CONPTY_POST_FULL_PAINT_SETTLE_MS;
+		if (until <= this.#postFullPaintSettleUntilMs) return;
+		this.#postFullPaintSettleUntilMs = until;
+		const hadPendingRender = this.#renderRequested || this.#renderTimer !== undefined;
+		// Reclaim any render that was queued during the in-flight composition:
+		// `#renderRequested` was set before the settle existed and would
+		// otherwise fire on the standard throttle inside the window.
+		this.#renderRequested = false;
+		if (this.#renderTimer) {
+			this.#renderTimer.cancel();
+			this.#renderTimer = undefined;
+		}
+		if (this.#postFullPaintSettleTimer) {
+			this.#postFullPaintSettleTimer.cancel();
+			this.#postFullPaintSettleTimer = undefined;
+		}
+		if (hadPendingRender) {
+			// Replay the absorbed request via the trailing settle timer so the
+			// caller's render still happens — just deferred to the end of the
+			// window. Subsequent `requestRender(false)` calls during the
+			// settle see this timer and fold into it (existing gate at L1263).
+			this.#postFullPaintSettleTimer = this.#renderScheduler.scheduleRender(() => {
+				this.#postFullPaintSettleTimer = undefined;
+				this.#postFullPaintSettleUntilMs = 0;
+				if (this.#stopped) return;
+				this.requestRender(false);
+			}, TUI.#CONPTY_POST_FULL_PAINT_SETTLE_MS);
+		}
+	}
+
+	#clearPostFullPaintSettle(): void {
+		if (this.#postFullPaintSettleTimer) {
+			this.#postFullPaintSettleTimer.cancel();
+			this.#postFullPaintSettleTimer = undefined;
+		}
+		this.#postFullPaintSettleUntilMs = 0;
+	}
 	#prepareForcedRender(clearScrollback: boolean): void {
 		const geometryChanged =
 			(this.#previousWidth > 0 && this.#previousWidth !== this.terminal.columns) ||
@@ -1927,6 +2031,7 @@ export class TUI extends Container {
 					clearViewport: true,
 					clearScrollback: !isMultiplexerSession(),
 				});
+				if (lines.length > height) this.#armPostFullPaintSettle();
 				this.#hasEverRendered = true;
 				return;
 			case "historyRebuild":
@@ -1935,6 +2040,7 @@ export class TUI extends Container {
 					clearViewport: true,
 					clearScrollback: !isMultiplexerSession(),
 				});
+				if (lines.length > height) this.#armPostFullPaintSettle();
 				return;
 			case "overlayRebuild":
 				this.#clearNativeScrollbackDirty();
@@ -1945,6 +2051,7 @@ export class TUI extends Container {
 					clearScrollback: !isMultiplexerSession(),
 				});
 				this.#emitViewportRepaint(lines, width, height, cursorPos);
+				if (baseLines.length > height) this.#armPostFullPaintSettle();
 				return;
 			case "liveRegionPinned":
 				this.#emitLiveRegionPinnedRepaint(