(terminal): switching to a hidden tab freezes the renderer for ~9 s

Switching to a hidden session tab freezes the renderer for about 8-10 s every time, whatever the amount of output replayed.

Measured 2026-10-10 in the live activity trace (v0.0.91; public/terminal-manager.js is identical on d14b035):

Reveal (UTC)	session	`writeChars` replayed	stall
23:56:07	6577a487	429 275	9.0 s
23:59:44	7615734f	26 118	8.5 s
00:03:05	6577a487	152 194	9.8 s

Every reveal in the window stalls, and the duration does not depend on volume (×16 in volume, same duration).
The main process stays alive meanwhile (osc.title events about every second); only the renderer is blocked.

Ruled out:

Replay write: xterm 6.0.0 parses 26 KB in 29 ms and 2 MB in 194 ms, measured under jsdom.
Scrollback reflow: resizing with 10k-50k scrollback lines takes 40-133 ms.
Click path JS: clearNotifications and the sidebar class.apply events land in the same millisecond.
Synchronous IPC: there is no sendSync.

Main suspect (plausible, not proven): the render path of showSession (~1291-1311).

suspendTerminalWebgl(previous) disposes the previous tab's WebGL addon.
restoreTerminalWebgl creates a new WebglAddon and context before the container gets .visible, i.e. on a display:none element.
forceRepaint then calls clearTextureAtlas and refresh.
atlasChanges and atlasCanvases stay 0, so this is not a repaint storm, but they do not measure context creation.

Instrument each step of showSession and its rAF callback under window.ATRACE as a reveal.timing event, so one tab switch shows which step takes the time.
Then fix:
- preferred: keep the WebGL context across switches, within the existing LRU cap;
- or load it after the container is visible, sized, one frame later.

Earlier, possibly related: xterm "task queue exceeded allotted deadline" overruns of 41-82 s on 2026-10-09 18:06-18:41. That warning means a single idle task overran its slot, so it is not explained by a chunked write. Its cause is not identified.

History check, v0.0.51 → v0.0.91 (read-only, git and the trace):

The switch path itself has not regressed. suspendTerminalWebgl(previous), restoreTerminalWebgl(next), replayHiddenBuffer and forceRepaint have been unchanged since fix(terminal): clear stale WebGL glyph atlas on reveal and on atlas rebuild #103/perf(terminal): suspend rendering work for hidden sessions in single view #115/perf(terminal): fully suspend xterm for hidden sessions (buffer-and-replay) #122 (v0.0.45-v0.0.51, mid-August). Electron 41.0.3, xterm 6.0.0, addon-webgl 0.19.0 and addon-fit 0.11.0 are identical from v0.0.51 to v0.0.91.
Only dependency change in the window: node-pty 1.1.0 → 1.2.0-beta.15 ((terminal): move node-pty to 1.2.0-beta.15 (#409) #431, v0.0.88, 2026-10-03), which bundles OpenConsole. Its cursor-position queries after a resize are a plausible source of stray input or wrong content. Nothing links it to the renderer stall.
More work around each switch:
- per-session panel visibility and width, so a switch can resize and send a ConPTY resize;
- the panel shell, which mounts or unmounts its own terminal and WebGL context ((panel): run a shell under the Changes list, in the session's own working directory #300, v0.0.81);
- the right-hand tool bar ((file-panel): move the session tools to a right-hand bar (#506) #511, v0.0.91).
Stall meter: render.stats.ms comes from a 1 s setTimeout. In .005, the stalls are 9042, 8460, 9841, 8637 and 9766 ms; earlier files have about 20 stalls of 3-17 s. atlasChanges stays 0 everywhere.
forceRepaint right after creation: clearTextureAtlas() runs right after the addon is created, which resets the atlas warm-up (93 idle glyph tasks, TextureAtlas.ts:122-146). The xterm "task queue exceeded allotted deadline" warnings come from that same IdleTaskQueue, probably as a symptom of the stall seen from inside the queue.

Discriminating checks:

switch.step marks in showSession / fitAndScroll, plus a PerformanceObserver on long-animation-frame.
An A/B without suspendTerminalWebgl(previous).
With the trace off, compare tags v0.0.87 / v0.0.88 / v0.0.91, and v0.0.51 as the environment control: if v0.0.51 also stalls, the cause is the environment (GPU driver, session count), not a regression.

test(terminal): pin that switching tabs keeps recent WebGL contexts a…

fix(terminal): keep recent WebGL contexts across tab switches and tim…

Merge main into wt/526, keep the #526 entry under Unreleased

Bisected live, same bench for every build: isolated test-pr, 5 plain terminals each streaming 30,000 lines, 15 switches, timed from showSession to two animation frames, window visible.

build	median switch
v0.0.80 / v0.0.85 / v0.0.88 / v0.0.90	67–79 ms
#508 head	70 ms
#511 head (#506, tools moved to a right-hand bar)	1963 ms
v0.0.91	2161 ms
main (v0.0.92)	1889–2192 ms
#532	37–42 ms

So the regression came in v0.0.91 with #511. The cost is one long animation frame of ~1.6 s per switch on main. #532 removes it, but the mechanism #511 added has not been identified: the WebGL dispose/recreate per switch already existed at 75 ms. Worth a look once #532 is in, in case something else in the same path is still expensive.

(terminal): keep recent WebGL contexts across tab switches and time e…

Navigation Menu

Repository navigation

(terminal): switching to a hidden tab freezes the renderer for ~9 s #526

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

(terminal): switching to a hidden tab freezes the renderer for ~9 s #526

Description

Activity

devsuitup commented on Oct 10, 2026

devsuitup commented on Oct 10, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions