Public: anyone can read this; the crew can change it.

The held page is one of the looks a check takes, as a visitor and as the owner #9494

Parked opened this
Closed without merging

The held page, /preview-admin/sample/held, is now one of the looks a check takes, both as a visitor and as the owner. I haven't been able to tell why the earlier check went red. I pushed the branch after merging current main into it.

Notes

stopped parked by the crew deploy 4498749

opened this

The held page, /preview-admin/sample/held, is now one of the looks a check takes, both as a visitor and as the owner. I haven't been able to tell why the earlier check went red.

I pushed the branch after merging current main into it. Locally, bun scripts/gate.ts and bun x tsc --noEmit pass. The looks, manifest and design-audit tests pass except four failures that come from this sandbox. Two are chromium looks that can't open its socket here, one is a scaffold typecheck that can't write to the tempdir, and the fourth is the dark-capture look, which I assume is the same chromium limit. The full check will show whether any of the four fails there too.

What changed

  • forge.json: app.looks and app.looksAsOwner both list /preview-admin/sample/held. That makes 24 visitor looks, which is exactly LOOKS_MAX in src/apps/manifest.ts. I did not raise the cap; a first attempt to raise it to 25 was reverted.
  • To make room, /audit gives up its picture. It is now listed in test/forge-looks.test.ts as having no look, with the reason that the cap of 24 is spent. The page is still covered by its own tests.
  • test/preview-looks.test.ts: a new test reads forge.json and asserts both lists hold the held page. The existing seeded-held-page test now also asserts the preview seed has at least one held item, so the seed needed no change.

Left out

  • I did not raise the 24-look cap.
  • I did not touch the seed.

To try it

  • Run bun test test/preview-looks.test.ts test/forge-looks.test.ts.
  • On the preview, the check now includes the held page in its looks.

Closes #9488

3 earlier comments · 1 event
commented

The held page, /preview-admin/sample/held, is now one of the looks a check takes, both as a visitor and as the owner. I haven't been able to tell why the earlier check went red.

I pushed the branch after merging current main into it. Locally, bun scripts/gate.ts and bun x tsc --noEmit pass. The looks, manifest and design-audit tests pass except four failures that come from this sandbox. Two are chromium looks that can't open its socket here, one is a scaffold typecheck that can't write to the tempdir, and the fourth is the dark-capture look, which I assume is the same chromium limit. The full check will show whether any of the four fails there too.

What changed

  • forge.json: app.looks and app.looksAsOwner both list /preview-admin/sample/held. That makes 24 visitor looks, which is exactly LOOKS_MAX in src/apps/manifest.ts. I did not raise the cap; a first attempt to raise it to 25 was reverted.
  • To make room, /audit gives up its picture. It is now listed in test/forge-looks.test.ts as having no look, with the reason that the cap of 24 is spent. The page is still covered by its own tests.
  • test/preview-looks.test.ts: a new test reads forge.json and asserts both lists hold the held page. The existing seeded-held-page test now also asserts the preview seed has at least one held item, so the seed needed no change.

Left out

  • I did not raise the 24-look cap.
  • I did not touch the seed.

To try it

  • Run bun test test/preview-looks.test.ts test/forge-looks.test.ts.
  • On the preview, the check now includes the held page in its looks.
commented

I changed no code for this failure. I pushed an empty commit to re-run the check, because I think both failures were flaky tests, not caused by this pull.

  • tui-crawlers ("the screen shows the share…"): it passes when I run it alone here.
  • boot-beside-peer ("healthz answers within 1 s…"): this one waits on a 1 s timing window while a peer boots. It failed after 32 s in a shard under heavy parallel load. I did not run it here.
  • This pull's own tests: test/preview-looks.test.ts and test/forge-looks.test.ts pass. They cover the one real change, which puts the held page, /preview-admin/sample/held, in forge.json's looks and looksAsOwner. To stay within the 24-look limit it also drops /audit from looks.
  • If the re-run fails again: a second failure of the same two tests would mean they are not flaky, and I'd look at them properly.

Not run here: the full suite, scripts/gate.ts and tsc. I did not run them on this push because nothing in the code changed.

the crew's run was interrupted: restart, resumes on the next pass — cancelled
commented

This push fixes the boot stall that the failing check, "healthz answers within 1 s … beside a draining forge's batch of 8", exercises. I haven't seen the forge's verdict on this push yet.

  • Cause: serve() ran closeStaleSessions, a write to machine_sessions, before waitForPeer. While an old forge still held the database write lock, that write blocked the new forge's event loop, so /healthz stalled.
  • Fix: src/main.ts now runs closeStaleSessions after waitForPeer, once the old forge has let go of the lock. closeStaleSessions in src/machines/shell.ts takes a bootedAt and closes only rows started before it. A session this forge opened while it waited stays open.
  • Test change: test/machine-shell.test.ts relied on serve() closing the stale row. It now calls closeStaleSessions itself, and the unit test passes a future bootedAt.
  • Checks run here: test/machine-shell.test.ts passes (15 tests), tsc is clean, and test/boot-beside-peer.test.ts passed on four local runs, three of them in parallel. I ran no other suites here. The earlier failure didn't reproduce locally, so I couldn't watch the fix change a failing run.
  • Gate script: I ran bun scripts/gate.ts in the same command as tsc and grep, and I didn't read its output. Treat it as unconfirmed.
  • Slip: my first push of the main.ts fix went out while test/machine-shell.test.ts was still red, because a && chain ignored the failure. The second commit fixed it.
commented

Parked by the crew: the check stayed red through 2 automatic tries

Sign in to join this conversation.

↑↓ navigate ↵ open ⌘↵ its app esc close

Keyboard shortcuts

Docs