From 7b22a2309c5f71b65dc1af123cf4d69c2d95d393 Mon Sep 17 00:00:00 2001
From: can1357 <me@can.ac>
Date: Sun, 16 Aug 2026 02:05:54 +0200
Subject: [PATCH] fix(tui): handled CPU-throttled loop stalls

---
 packages/tui/CHANGELOG.md               |  3 +++
 packages/tui/src/loop-watchdog.ts       | 17 +++++++++--------
 packages/tui/test/loop-watchdog.test.ts |  5 +++--
 3 files changed, 15 insertions(+), 10 deletions(-)

diff --git a/packages/tui/CHANGELOG.md b/packages/tui/CHANGELOG.md
index a67c956de..17050627d 100644
--- a/packages/tui/CHANGELOG.md
+++ b/packages/tui/CHANGELOG.md
@@ -32,6 +32,9 @@
 ### Fixed
 
 - Fixed inline images rendering permanently cropped on Kitty direct-placement terminals (WezTerm, Warp) when an image block straddled the viewport top during streaming: placements are now clipped to the visible slice at write time, and a placement id whose cells reached native scrollback is never re-used ([#8070](https://github.com/can1357/oh-my-pi/pull/8070) by [@voonfoo](https://github.com/voonfoo))
+### Fixed
+
+- Fixed long CPU-bound event-loop stalls being misclassified as system sleep and omitted from `ui.loop-blocked` diagnostics ([#8442](https://github.com/can1357/oh-my-pi/pull/8442) by [@Mustaqeem66](https://github.com/Mustaqeem66)).
 
 ## [17.2.12] - 2026-08-08
 
diff --git a/packages/tui/src/loop-watchdog.ts b/packages/tui/src/loop-watchdog.ts
index 10f39f00b..79264374c 100644
--- a/packages/tui/src/loop-watchdog.ts
+++ b/packages/tui/src/loop-watchdog.ts
@@ -6,7 +6,7 @@ export interface LoopWatchdogOptions {
 	intervalMs?: number;
 	/** A tick later than this past its deadline counts as a block. Default 250. */
 	thresholdMs?: number;
-	/** Overshoot beyond this is suppressed only when the process also burned no CPU. Default 60_000. */
+	/** Overshoot beyond this is suppressed only when the process burned negligible CPU. Default 60_000. */
 	sleepMs?: number;
 	/** Monotonic clock source; injectable for tests. Default `performance.now`. */
 	now?: () => number;
@@ -26,13 +26,13 @@ interface LoopWatchdogTimer {
 }
 
 /**
- * Fraction of a missed interval that must show up as CPU time for the overshoot
- * to count as a synchronous stall rather than system sleep. A suspended process
- * resumes having burned essentially nothing; a wedged one burned the interval on
- * a core. Half leaves room for a gap that is partly sleep and partly work, which
- * is still worth reporting.
+ * Fraction of a missed interval that may be process CPU time while the gap is
+ * still treated as system sleep. Keep this near zero: cgroup throttling and
+ * scheduler contention can make a CPU-bound loop consume far less CPU than wall
+ * time. One percent allows a little measurement/background jitter while erring
+ * toward reporting a severe stall instead of hiding it.
  */
-const CPU_BUSY_RATIO = 0.5;
+const CPU_BUSY_RATIO = 0.01;
 
 /**
  * Always-on event-loop lag probe. Each tick is scheduled `intervalMs` ahead of
@@ -51,7 +51,8 @@ const CPU_BUSY_RATIO = 0.5;
  * A long overshoot is classified by CPU time rather than by duration. System
  * sleep and a CPU-bound wedge both produce an arbitrarily large gap, so duration
  * alone cannot tell them apart, and suppressing on duration discards exactly the
- * worst stalls. Only a gap the process did not spend CPU on is treated as sleep.
+ * worst stalls. Only a gap the process spent negligible CPU on is treated as
+ * sleep. CPU accounting is process-wide, so worker activity errs toward logging.
  */
 export class LoopWatchdog {
 	#intervalMs: number;
diff --git a/packages/tui/test/loop-watchdog.test.ts b/packages/tui/test/loop-watchdog.test.ts
index db3333ce7..1fe07e59d 100644
--- a/packages/tui/test/loop-watchdog.test.ts
+++ b/packages/tui/test/loop-watchdog.test.ts
@@ -288,12 +288,13 @@ describe("LoopWatchdog long-block classification", () => {
 		expect(warnSpy).not.toHaveBeenCalled();
 	});
 
-	test("reports a long block that spent most, but not all, of the gap on CPU", () => {
+	test("reports a long CPU-bound block when the process is CPU throttled", () => {
 		const warnSpy = vi.spyOn(logger, "warn").mockImplementation(() => {});
 		const h = cpuHarness();
 
 		h.wd.start();
-		h.set(120_250, 90_000);
+		// Ten percent CPU is still real work, not suspension.
+		h.set(82_641, 8_200);
 		h.fireTick();
 
 		expect(warnSpy).toHaveBeenCalledTimes(1);