Repository navigation
finding(e2e): the boot-continuity probe cannot fail for any gate that decides after a session exists #6569
Description
Found while implementing #6507 (PR #6565). Recording the measurement, not proposing an implementation.
What is missing
e2e/console-boot-indicator.spec.ts carries the campaign's acceptance bar: after React's first commit, the viewport centre must never stop being covered. It catches the objectui#6378 defect on the three gates #6506 converted, and it passes on a signed-OUT boot.
It cannot be made to fail for any gate that decides after a session exists. #6507 converted seven such gates. Per-site scenarios were built for them and they do not bind.
The measurement
Environment: production apps/console bundle, vite preview, Playwright driving the pre-installed Chromium 141.0.7390.37 at PLAYWRIGHT_BROWSERS_PATH=/opt/pw-browsers via executablePath. The harness runs; the acceptance spec passes 4/4.
Signed-in mock scenarios were built for RequireAiSurface (a runtime serving no agent) and AppContent's no-accessible-app bounce. Both drive cleanly, reach their destinations, and pass. Then:
- Ablated the fix — replaced all seven
RedirectWithSplashelements with a bareNavigatein source, rebuilt.packages/app-shell/dist/console/ConsoleShell.jsandAppContent.jsconfirmed rebuilt; 508 console dist assets fresh. Both scenarios still passed. - Added 20x CPU throttling (
Emulation.setCPUThrottlingRate), since objectui#6378 measured its window on a throttled profile. Both scenarios still passed.
So the scenarios were deleted rather than committed: an assertion that cannot fail is not coverage.
The diagnosis
Not a missing browser, and not a missing endpoint. The probe treats the pre-React #boot-splash as covering — correctly, it is on screen. On these mocked boots the redirect chain resolves before that indicator is torn down, so at the moment the gate decides there is no blank for the sampler to land in.
A per-site gate that CAN fail therefore needs the window reproduced with the indicator already gone: a slower boot to the decision point, or a fixture that tears the indicator down on React's first commit the way a real cold load does, or a probe that starts sampling from a later anchor.
What is NOT established
- Whether the window is unreachable in principle on these boots, or only under this fixture's timing. One fixture, two sites.
- Any rate. Each scenario was run a handful of times, not sampled.
- Whether the same is true for the other five converted gates; two of them (
RootRedirect,AuthenticatedRoute) are not mounted byapps/consoleat any session state and cannot be reached by this bundle at all.
Reusable material
The endpoint set for a signed-in mock boot was read off a real boot rather than guessed: /api/v1/auth/get-session, /auth/config, /auth/organization/list, /auth/me/localization, /api/v1/ai/agents, /api/v1/meta/{app,object,view,page,report,dashboard,_drafts}, /api/v1/discovery, /api/v1/runtime/config, /api/v1/i18n/{locales,translations}, /api/v1/data/sys_*, /api/v1/approvals/requests.
Standing coverage for the seven gates meanwhile is at the DOM level, and it does bind — 7 of 12 red under the same ablation with all five control cases green: packages/app-shell/src/console/__tests__/bootRedirectCoverage.test.tsx and AppContent.bootRedirectCoverage.test.tsx.
Generated by Claude Code
Activity
os-support-ai commentedon Aug 26, 2026
Triage: graded — closing not planned. This is an honest, well-measured record of a fixture limit, and the record is the value; the card itself proposes no implementation.
Rationale on the four axes, briefly: standing coverage for the seven post-session gates already binds at the DOM level (7/12 red under ablation with green controls —
bootRedirectCoveragesuites), so no invariant is unguarded today; hardening the e2e probe (slower-boot fixture, later sampling anchor, or teardown-on-first-commit) is test-infra capability expansion with no current pull — startup-focus says don't build it speculatively; and the ablation methodology here (delete assertions that cannot fail) is exactly right and needs no further action.Reopen is free if a real blank-frame regression on a post-session gate ever ships past the DOM-level coverage — that event would supply the pull this card currently lacks. The endpoint set under "Reusable material" stays findable here for whoever builds a signed-in mock boot next. Maintainer veto by reopening, as ever.
Generated by Claude Code