A test that chdir()s into a temp dir and deletes it cannot exist on Windows: the process's own cwd is locked
env paths · case
Symptom
Section titled “Symptom”A regression test for “the CLI survives being launched from a deleted directory” passes on macOS and Linux and fails on Windows in its SETUP, before the assertion:
error: EBUSY: resource busy or locked, rm 'C:\Users\user\AppData\Local\Temp\ocx-unlinked-cwd-3epkdD' at removeTreeWithRetry (tests/helpers/remove-tree.ts:28:83)(fail) cli wiring > interactiveGuardOk safely evaluates without throwing when cwd is unlinked [2611.94ms]The 2.6 s is the retry helper spending its whole 50 × 50 ms budget on an operation that can never succeed, then rethrowing.
import { mkdtempSync, rmSync } from "node:fs";import { tmpdir } from "node:os";import { join } from "node:path";
const dir = mkdtempSync(join(tmpdir(), "cwd-lock-"));process.chdir(dir);rmSync(dir, { recursive: true, force: true }); // POSIX: fine. Windows: EBUSY, every time.console.log(process.cwd()); // POSIX prints the dead path; never reached on WindowsPOSIX unlinks a NAME. A process standing in a directory keeps a valid handle to the inode
after the name is gone, so rmdir succeeds and getcwd() starts failing with ENOENT — which is
exactly the state the test wants to exercise.
Windows holds the current directory as an open handle without FILE_SHARE_DELETE for as long
as it IS the current directory. Deleting it is refused with ERROR_SHARING_VIOLATION
(EBUSY) until some process chdirs elsewhere. There is no “directory exists but has no
name” state to enter, so the precondition is unreachable, not merely slow.
This is a different failure class from an unowned child holding a handle
(async-child-holds-dir-after-stop): the holder here is the test process itself, and no amount
of waiting or reaping changes that.
Workaround
Section titled “Workaround”Skip the case on Windows, with the platform fact in the predicate’s comment so it reads as a boundary rather than a muted failure:
// Windows locks a process's cwd: it cannot be unlinked, so the state under test// cannot exist there. The cwd-healing code path is covered separately.test.skipIf(process.platform === "win32")("... when cwd is unlinked", () => { /* unchanged */ });Two tempting alternatives are wrong:
chdirback to the original directory before deleting. The delete now succeeds and the test runsinteractiveGuardOk()from a perfectly valid cwd, asserting nothing.- Inventing a Windows-reachable “bad cwd” (a revoked ACL, a dropped drive mapping). That exercises a different failure mode dressed up to keep the checkmark green.
A skip is honest here because there is no coverage to lose: the platform cannot enter the state. If Windows coverage of the guard’s throw branch is wanted, it belongs in a separate test that injects the throw directly.