Builds and tests leave craft-* temp directories behind

On 2026-10-14 the shared Mac ran out of disk twice (248 MB, then 1.1 GB free), killing simulator runs. One cause was 1,983 stale craft-* directories older than an hour in $(getconf DARWIN_USER_TEMP_DIR), which another session deleted by hand to free ~13 GB (with ~/.bun/install/cache).

Something in Craft (the build pipeline, native screen compilation, template/fixture tests, or the E2E harness) creates a craft-* temp dir per run and never removes it, including on success.

Wanted:

• Find every place that creates a craft-* temp dir (mkdtemp or similar) and remove it when the run ends, on success, failure and interrupt (try/finally plus a signal handler where a process can be killed mid-run).
• Long-lived dirs that are meant as caches should live under a single, bounded, named cache dir with pruning, not in temp.
• A test that runs a representative build/test and asserts no craft-* dir is left behind.
• Note it in the release notes so apps' builds stop leaking after upgrading.

Fixed in 8e48f41, released in v0.0.187.

Every Craft scratch directory now comes from packages/typescript/src/temp-dir.ts. A directory is removed in three ways:

1. by the caller when it is done (withTempDir, or removeTempDir in a finally);
2. on process exit and on SIGINT/SIGTERM/SIGHUP. Node skips exit listeners on a signal's default action, so the signals get their own handlers. When no other listener is registered, the process then exits with 128+n, as the default action would;
3. by a sweep, for runs killed with SIGKILL. Names carry the owning pid (craft-<label>-p<pid>-XXXXXX), and a later process's first createTempDir removes dirs whose pid is dead. Only that naming is ever swept, dirs younger than 60 s are skipped, and a dir marked with keepTempDir() (.craft-keep) is left alone.

Leaks found and fixed

Checked and clean: the init()/build() paths Stacks' build:ios/build:android call (no temp dirs), xcodegen/gradle generation (writes into the caller's output), Zig tests (.zig-cache/tmp via std.testing.tmpDir, plus fixed /tmp file names), the Windows multi-window verifier, and CI's $RUNNER_TEMP use. No temp dir was serving as a cache. The updater's craft-updates/<app> download dir is one fixed path per app, so it is already bounded, and it is removed after install.

Tests

• packages/typescript/src/__tests__/temp-dir.test.ts: naming, removal, withTempDir on throw, signal listeners only while dirs are live, sweep (removes dead-pid dirs only; keeps live, young, kept, foreign and legacy names).

• packages/ios/src/temp-dir-leaks.test.ts: each case runs in a child process with TMPDIR set to an empty sandbox, and asserts no craft-* is left after:

  • a Stacks-style iOS + Android init + build that succeeds;
  • the same build failing with an uncaught error;
  • SIGINT and SIGTERM mid-run (exit 130/143);
  • SIGKILL, followed by the next process's sweep;
  • the whole iOS builder suite plus a deliberately failing test file.

  With the exit/signal hooks disabled, 4 of the 6 cases fail.

Locally, after the fix: ios (250), android (134), SDK (668) and the 17 touched scripts/*.test.ts suites all pass, with zero new craft-* dirs in the real temp folder.

Apps stop leaking once they pick up craft-native ≥ 0.0.187. Known limit: if a run is blocked in Bun.spawnSync (xcodebuild) and only the parent gets SIGTERM, cleanup waits for the child to finish. If the run is SIGKILLed first, the next run's sweep removes the dir.