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

Full suite red: test/boot-production.test.ts #12659

Merged opened this in answer to

Check · Review · Land · Live done — Landed and live in production.

Merged

test/boot-production.test.ts now holds each production boot for 10 s instead of 20 s, so the file should fit the 20 s per-file budget. I couldn't run it to completion here: the sandbox can't bind the unix socket the booted forge listens on, so five of its six tests fail (EPERM).

Notes

shipped production is serving it deploy e5f3b40

opened this

test/boot-production.test.ts now holds each production boot for 10 s instead of 20 s, so the file should fit the 20 s per-file budget. I couldn't run it to completion here: the sandbox can't bind the unix socket the booted forge listens on, so five of its six tests fail (EPERM). The file fails the same way without my change, so the forge's check is where this gets confirmed.

  • Why it was slow: the three boots (FORGE_ENGINE inline, unset, worker) already run in parallel, so the file's time was the hold plus setup and shutdown. That came to 21.0 s, just over the budget.
  • The change: HOLD_MS goes from 20 s to 10 s, and the test names and header comment say "ten seconds". docs/ARCHITECTURE.md (the line saying the boot is held up for twenty seconds) is updated to match.
  • What stays the same: the assertions are unchanged. Each boot must still be alive and answering /healthz at the end of the hold, with no uncaught exception, no DataCloneError, and the same expected log lines.
  • What I left out: I did not add a SLOW entry in scripts/gate.ts, since making it faster was possible. I also did not rework the boots to share a process.
  • Trade-off: the guard now watches a boot for half as long. The 0.7.599 failure showed up seconds after /healthz first answered, so 10 s still covers it.
  • Checks: bun x tsc --noEmit and bun scripts/gate.ts both pass.
  • To try it: run bun test ./test/boot-production.test.ts somewhere that can bind a unix socket (or sets FORGE_TEST_SOCKET for the child). It should finish in about 11 s.

Closes #12642

commented

test/boot-production.test.ts now holds each production boot for 10 s instead of 20 s, so the file should fit the 20 s per-file budget. I couldn't run it to completion here: the sandbox can't bind the unix socket the booted forge listens on, so five of its six tests fail (EPERM). The file fails the same way without my change, so the forge's check is where this gets confirmed.

  • Why it was slow: the three boots (FORGE_ENGINE inline, unset, worker) already run in parallel, so the file's time was the hold plus setup and shutdown. That came to 21.0 s, just over the budget.
  • The change: HOLD_MS goes from 20 s to 10 s, and the test names and header comment say "ten seconds". docs/ARCHITECTURE.md (the line saying the boot is held up for twenty seconds) is updated to match.
  • What stays the same: the assertions are unchanged. Each boot must still be alive and answering /healthz at the end of the hold, with no uncaught exception, no DataCloneError, and the same expected log lines.
  • What I left out: I did not add a SLOW entry in scripts/gate.ts, since making it faster was possible. I also did not rework the boots to share a process.
  • Trade-off: the guard now watches a boot for half as long. The 0.7.599 failure showed up seconds after /healthz first answered, so 10 s still covers it.
  • Checks: bun x tsc --noEmit and bun scripts/gate.ts both pass.
  • To try it: run bun test ./test/boot-production.test.ts somewhere that can bind a unix socket (or sets FORGE_TEST_SOCKET for the child). It should finish in about 11 s.
Approved

Approved. A single-purpose change that speeds up a slow test, which fits the pipeline direction (#1242, a time budget per file) and keeps the main-green goal. The assertions are untouched, and the test names, header comment and ARCHITECTURE.md line all agree on ten seconds. There is no user-facing surface, no new dependency, no new input, and no README or TUI change needed. The pull is green. The preview is the unchanged app, since only a test and a doc changed.

Two things I would do next:

  1. Derive the hold from a single named constant in the test names and the comment, or drop the number from the prose, so the next tuning does not touch three places plus the docs.
  2. Watch the forge check for the file time. 10 s of hold plus setup was estimated at about 11 s, but it was not run here (no unix socket in the sandbox). If it comes in near 20 s under load, share one process across the boots instead of cutting the hold again.
Sign in to join this conversation.

↑↓ navigate ↵ open ⌘↵ its app esc close

Keyboard shortcuts

Docs