Main-process exceptions no longer block quitting (D28): the check, the strip's fade timer, two test races by siyabendoezdemir · Pull Request #11 · CIYAhq/orbit-browser

siyabendoezdemir · 2026-10-10T00:23:53Z

Was stacked on PR #10 (base lead/polish2), since both change tests/e2e/helpers.ts. #10 has merged, so the base is main now.

What

Product. If Orbit's main process hits an exception nothing catches, Electron's default shows a modal error box. The box holds the main thread until someone dismisses it, so Orbit can't quit. A user would see a stack trace in a box, then an app that won't quit. Now the exception is logged ([orbit] uncaught exception: …), and Orbit carries on (src/main/index.ts).
E2E harness (tests/e2e/helpers.ts). app.close() now checks the quit:
- a test fails, showing the stack, if its main process logged such an exception, including one thrown while quitting. orbit.takeUncaught() claims one thrown on purpose;
- if Orbit still hasn't quit 15 s after being asked, the test fails and Orbit is killed. One stuck quit then fails one test instead of hanging the whole run.
The exception behind the SV-03 hangs, found by the check and fixed (13a2494, src/main/split/strip.ts):
- the split's pane strip fades out for 120 ms before its view hides. If its window closed during the fade, which happens when a test quits right after the strip hid, the fade's timer still ran;
- by then view.webContents was undefined, so reading isDestroyed() threw TypeError: Cannot read properties of undefined (reading 'isDestroyed');
- now the window's closed handler clears the strip's show, hide and fade timers, and every check of the strip's view handles a view whose page is gone.
Decision D28. docs/ARCHITECTURE.md §9 explains the check.
Two test races from the full run with the check (no product change):
- MW-10 (5efc435): fromMini sends a command from the Mini Window's own page. A close can shut that page before Playwright gets the evaluate's reply ("Target page, context or browser has been closed"). The error is now ignored, but only when the page really closed.
- SV-17 (721fdc8, cherry-picked from the chrome lane's 442766d): a pane added a moment ago grows in from zero width (SV-26), so it may have no box yet. The views-over-panes poll threw on that and failed at once; now it retries.

Why

Since 20:39 UTC, five runs ended with a test hanging in its afterEach's app.close(), then "Worker teardown timeout of 60000ms exceeded". That fails the job even when the test passes on retry:
- macOS CI on PR Split rest holds its row (W2), a Favorite split is one Favorite (D19), Spaces dialog fade #7 (CB-21), and on main at cc43e90 (MW-14);
- local full runs: SV-03 twice and TB-01 twice (one of those in the media verifier's run).
Playwright's close evaluates app.quit() through the inspector, then waits for the process to exit. No Orbit handler cancels a quit.
Reproduced: a main-process setTimeout that throws makes app.close() hang, with the main thread parked in do_poll (the box's nested loop). A rejected promise doesn't.
With the check in place, 15 repeats each of SV-03 and TB-01 logged one exception, in SV-03: the strip's fade timer above. That explains the SV-03 hangs. The TB-01 and macOS hangs (CB-21, MW-14) haven't shown theirs yet; when they do, the check names the test and the stack instead of hanging.

Tests

New: [TB-22] an exception nothing caught in the main process is logged; Orbit still quits and restores (D28) (core.spec.ts). It throws in a main-process timer, checks the log, quits and relaunches with the same profile.
- Without the fix it fails. Nothing is logged, and the quit hangs until the harness kills Orbit at 15 s: still running (main thread: do_poll.constprop.0).
- With the fix it passes.
New: [SV-12] a window that closes while its pane strip fades out closes cleanly (D28) (split.spec.ts). It opens a second window so Orbit keeps running, starts the strip's fade, closes the strip's window, and waits past the fade.
- Without the fix it fails 3/3 with the TypeError above. With the fix, it and the three other strip tests passed 20/20 (5 runs each).
Scratch checks, since deleted:
- a quit that never happens (before-quit prevented) is killed at 15 s, and its test fails;
- a clean quit passes and takes 100–350 ms. 40 launch, split and close cycles under load had no hang.
Full suite at 4fe445a (1440×900, 1 worker, 11.9 min): 216 passed, 2 failed. The failures were MW-10 and SV-17, the two test races above. No test logged an uncaught exception, and no quit needed the kill. After the fixes, MW-10 and SV-17 passed 12/12 (6 runs each).
With the strip fix, 20 repeats each of SV-03 and both TB-01 tests: 60/60 passed, with no exception logged.
Full suite at 13a2494 (1440×900, 1 worker, 9.4 min): 219/219 passed. No exception logged and no kill.
npm run typecheck and npm run lint exit 0; unit tests 252/252.

Review

Independent review (reviews/lead-quit.md in the Orca context, at 13a2494): APPROVE WITH FIXES, 1 must-fix and 6 follow-ups.
- It reproduced the root cause with the real bug. On the base build, the strip's fade timer firing after its window closed hung 5 of 8 quits, each with the main thread in Electron's error box. With this PR, all 16 such quits took 111–390 ms.
- Its full suite at 1440×900 passed 218 of 219; the one failure was CB-22's load-sensitive budget.
- The TB-01 and macOS hangs (CB-21, MW-14) are still unexplained. 102 repeats of those tests and 76 quits with timers pending found no other exception.
Lead fix (D18), 504872a (test harness only): the quit check missed an exception whose log ran past about 4,000 characters, because it trimmed stderr before looking for the marker. It now looks first and keeps only a marker-length tail otherwise. A new TB-22 test throws a 5,000-character message: it fails without the fix (takeUncaught() returned '') and passes with it, 3/3. core + split + smoke e2e: 62/62.
7eaec40 merges PR Lead polish 2: full-split drop target, Favorite splits past 12, e2e evaluate race #10's two review fixes into this branch; the reviewer checked them at d2e4b8d.
d19b9da, CI only (D18): the two Linux jobs are pinned to ubuntu-24.04. GitHub moves ubuntu-latest to Ubuntu 26 from 2026-10-19, and the e2e job's apt packages are only known to work on 24.04, the image ubuntu-latest runs today.
The follow-ups are tracked in the lead's STATUS.md:
- start the 15 s clock when the quit reaches main;
- check a quit that happens on its own;
- log unhandled promise rejections the same way;
- take a sample of the stuck process on macOS before the kill;
- write the log to a file in packaged builds;
- guard Peek's settle timer like the strip's.

Local check (merge gate, D42)

GitHub Actions is off for cost, so this local check is the merge gate. On d19b9da, in this PR's worktree (Linux, Xvfb, one worker):

npm run build; npm run typecheck; npm run lint; npx vitest run
xvfb-run -a -s "-screen 0 1440x900x24" ./node_modules/.bin/playwright test --workers 1 --retries 0

Build, typecheck and lint exit 0; unit 255/255 (22 files).
E2E: 220 passed, 1 failed (10.7 min, load about 30). The failure was SV-13's mid-drag check (split.spec.ts:115): the pane's view reached 527 px against more than 533 while the machine was loaded. Rerun alone, it passed (4.7 s), as in earlier runs. This PR's only split change is the pane strip's timers (strip.ts), not the divider drag. SV-03, whose quit hung in PR Lead polish 2: full-split drop target, Favorite splits past 12, e2e evaluate race #10's run without this fix, passed.
No test logged [orbit] uncaught exception:, and no quit needed the kill. Since this head contains PR Lead polish 2: full-split drop target, Favorite splits past 12, e2e evaluate race #10, this run is also Lead polish 2: full-split drop target, Favorite splits past 12, e2e evaluate race #10's check.
macOS: not run, so unverified. That covers the macOS hangs this PR targets (CB-21 on PR Split rest holds its row (W2), a Favorite split is one Favorite (D19), Spaces dialog fade #7's CI, MW-14 on main): the fix is Electron's error box, which is the same code on macOS, but no macOS run has confirmed them gone.
d19b9da pins the Linux CI jobs to ubuntu-24.04. The workflow is disabled now (D42), so that change is inert; it keeps the file right if the workflow is ever turned back on.

🤖 Generated with Claude Code

Note

Medium Risk
Changes main-process error handling and split strip lifecycle timers; behavior on real bugs shifts from a blocking dialog to continued execution, which is intentional but affects how failures surface in production.

Overview
Uncaught main-process exceptions no longer block quit: Electron’s modal is replaced by logging [orbit] uncaught exception: in src/main/index.ts (D28).

Split pane strip (strip.ts) clears show/hide/fade timers when the window closes and uses a live() guard so fade timers don’t touch destroyed views—the failure mode that hung app.close() in E2E.

E2E harness (helpers.ts): launchOrbit wraps app.close() to fail on any uncaught main exception (with takeUncaught() for intentional throws), kill the app after 15s if quit stalls, and detect the marker even when stderr writes exceed ~4k chars. New TB-22 / SV-12 tests cover logging, quit/restore, long logs, and strip fade + window close. MW-10 ignores evaluate errors when the Mini Window page closes first; SV-17 polling tolerates panes still growing in (SV-26).

CI pins Linux jobs to ubuntu-24.04 instead of ubuntu-latest; ARCHITECTURE.md §9 documents the harness behavior.

^{Reviewed by Cursor Bugbot for commit d19b9da. Bugbot is set up for automated code reviews on this repo. Configure here.}

…blocked quitting (D28) Electron's default uncaughtException handler shows a modal error box that holds the main thread until someone dismisses it, so Orbit couldn't quit. Under Xvfb nobody dismisses it: e2e app.close() hung, then Playwright's worker teardown, which failed whole jobs (macOS CI on PR #7 and main, three local full runs). A user would have seen the box and an app that wouldn't quit. - src/main/index.ts: log "[orbit] uncaught exception:" and carry on. - tests/e2e/helpers.ts: app.close() fails a test whose main process logged one (takeUncaught() claims one thrown on purpose) and kills an Orbit that still hasn't quit 15 s after being asked, so one stuck quit can't hang the run. - core.spec.ts: [TB-22] a throw in a main-process timer is logged, Orbit quits and the session restores. Without the fix nothing is logged and the quit hangs until the harness kills it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… before the evaluate replies fromMini sends a command from the Mini Window's own page. mini.control's close can close that page before Playwright gets the evaluate's reply, which then fails with "Target page, context or browser has been closed" (MW-10, once in a full run). The helper now ignores that error only when the page really closed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… growing in A pane added a moment ago springs open from zero width (SV-26), so its first bounding box can be empty. The poll now retries instead of failing at once. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

cursor · 2026-10-10T00:24:04Z

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.