Public: anyone can read this; the crew can change it.
Full suite red: test/check-install.test.ts #12897
Check · Review · Land · Live done — Landed and live in production.
The bun installs from a filled cache that is read-only test in test/check-install.test.ts now sets BUNTMPDIR to the cell's own writable home. The file passes (27 tests) and bun x tsc --noEmit is clean. Cause.
shipped production is serving it deploy e5f3b40
Nothing needed fixing. The "failure" in the deploy report isn't a test failure. The preview is up, and /healthz answers 200 at the preview URL. The log you pasted shows the forge's own process lines, ending in a SIGTERM drain of the forge itself. That is the forge restarting, not this app failing to boot. "No tests declared, none run" is the preview deploy's note: forge.json has no app.test, and the real test run is the check.
The pull's one change is in test/check-install.test.ts. The read-only-cache install in prefillInstallCache (#7543) now gets BUN_TMPDIR: home. Bun then stages in a tmp directory the test can write to, not the ambient one. That is a fix to the test's environment; the code under test is unchanged. I ran that file here and all 27 tests pass.
I changed nothing and pushed nothing. I did not run gate.ts, tsc or the full suite, since the commit touches only a test env var and the forge runs the full check on push.
The test/check-install.test.ts fix was already on this branch and pushed, and your message says the full suite is green again (check run 01m7dw4e73tqcsd0qj1ysvf2r5, c6641d1). I made no new changes. I re-ran the file here and all 27 tests pass.
- What changed: commit 97b442020 gives the read-only-cache install its own
BUN_TMPDIR. This was the cause of the red suite. - What I left alone: the test's assertions are unchanged. I did not touch anything else.
- How to try it: run
bun test test/check-install.test.ts.
I did not run the type check or the gate script, because there was nothing new to push.
One-line test-env fix: the read-only-cache install in prefillInstallCache (#7543) now stages in a writable BUN_TMPDIR, so the full suite goes green again. This serves the pipeline-first direction (main green). The code under test is unchanged and no assertions were weakened. It touches no src/web, so the experience checks do not apply. Two things I would do next: 1) split the long inline comment on the env line into its own line above it so the three reasons read separately; 2) find out why the ambient tmp is unwritable in the check cell, and set BUN_TMPDIR once in the shared test env helper (envFor) so other install tests cannot hit the same flake.
The
bun installs from a filled cache that is read-onlytest intest/check-install.test.tsnow setsBUN_TMPDIRto the cell's own writable home. The file passes (27 tests) andbun x tsc --noEmitis clean.Cause. The test spawns
bun installwith a stripped environment (onlyPATH,HOMEand the registry URL). Bun then looks for a staging directory in the ambient temp location. When the full suite runs four files at once, that location wasn't writable, so bun exited 1 withEACCES accessing temporary directory. This was a flaw in the test, not incellInstall.Not reproduced. The failure didn't occur in this sandbox, even before the change. I haven't run
--rerun-each 10, so I can't show the flake is gone; the forge's full check on this push is the real confirmation.Left alone. Production code is unchanged. The cell command in
src/checks/install.tsalready points the cache at a copy under$HOME, so it's unaffected. If a real cell can lack a writable temp dir, that is a separate question.I did not run
scripts/gate.tsor the design lint; the change touches only that one test line.Closes #12882