Detached Child

Container-green, host-red: a detached child holds the test's tempdir

Problem

A factory/CI gate runs a hook's self-test inside a Linux container and reports green. The hook is then wired into settings.json (and its self-test into the session receiver checks) on a Windows host. On that host the same suite exits 1 part-way through: the first block of checks prints ok, then a PermissionError [WinError 32] traceback from tempfile.TemporaryDirectory.__exit__, and every later check never runs. The receiver now rejects the next session packet on a "trusted check" that never reached its second case.

Observed 2026-08-29 in a private steering repository (board-drift sweep hooks): 19/19 gate checks green in-container; on the host 15/57, then abort.

Context / Trigger Conditions

• Windows host; Python 3.12/3.13; suite uses tempfile.TemporaryDirectory() per test.
• The code under test launches a long-lived detached child: subprocess.Popen([...], start_new_session=True) with no cwd= argument.
• Some test runs the parent with subprocess.run(..., cwd=td) where td is the tempdir.
• Symptom order: early ok lines → WinError 32 ... being used by another process: 'C:\\Users\\...\\Temp\\tmpXXXX' naming the directory, not a file → suite aborts.
• Red herring: the first tempdir you suspect (the one whose test polls for an output file) is not the holder. Adding ignore_cleanup_errors=True there changes nothing.

Solution

1. Find the holder by the frame, not the message. The traceback's in test_... frame names the test whose tempdir is locked. Look for cwd=td in that test.

2. Fix the product, not the test: the detached launcher pins its child's cwd to the project root (or another directory that outlives the test):

   root = os.environ.get("CLAUDE_PROJECT_DIR")
   if not root or not os.path.isdir(root):
       root = os.path.dirname(os.path.dirname(_HERE))
   subprocess.Popen([sys.executable, _DETACHED], cwd=root,
                    stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
                    start_new_session=True)
   

   An inherited cwd is a latent defect in production too: the child keeps whatever directory the hook was launched from locked for its lifetime.

3. Publish outputs atomically in the child: write <path>.tmp, then os.replace(tmp, path). A poller that keys on path.exists() then never sees a half-written or still-open file, which is the second Windows-only lock.

4. Compare path parts, not joined suffixes: Path(p).parts[-3:] == (".claude", "state", "x.json") instead of p.endswith(".claude/state/x.json").

5. On the test whose child may still be exiting at block end, TemporaryDirectory( ignore_cleanup_errors=True) is a legitimate belt-and-braces, but it is not the fix.

Verification

Run the suite on the host: exit 0 and the full check count (here all green (57 checks)). Before the fix: exit 1, the early block only, WinError 32 naming a temp directory. Confirm with grep -n "cwd=td" that the located test is the one in the traceback frame.

Example

Three defects survived a 19-check container gate and were found only by running the self-test on the host before merging: inherited cwd (abort), non-atomic cache write (race), slash-joined path assertion (backslashes). Fixed forward on the factory branch (945c684), re-run 57/57, then merged. Rule adopted in the project: a PR that wires a hook into settings.json or a receiver check into session-boundary.json gets its self-test run on the host before merge — the gate is not the host.

Notes

• Windows refuses rmdir on a directory that is any process's cwd; Linux does not, which is why the container never sees it.
• start_new_session=True is a POSIX setsid; on Windows it is ignored, so "detached" is weaker there too — the child survives parent exit only because Windows does not kill children by default. Measure, do not assume.
• See also: windows-claude-code-env (Problem 9, CRLF; cp1252 console encoding) for the other Windows-only classes that pass a Linux gate.

References

• Python tempfile.TemporaryDirectory(ignore_cleanup_errors=...) — added in 3.10; it swallows the error, it does not release the holder.
• Python subprocess.Popen — cwd and start_new_session semantics per platform.