Verification: the suite writes into the checked-in .the-loop/ when run from the repo root

The verification node's captured output, one section per row of testing-plan.md. Environment: Python 3.11, uv 0.12.17, uv sync --all-packages against the committed uv.lock, no ~/.the-loop/.

Results

Row	Command	Outcome
T1	uv run --project cli python -m pytest -q cli (repo root), then both trees' status	4528 passed, 1 skipped; both trees untouched, tracked and ignored
T2	cd cli && uv run python -m pytest -q, then both trees' status	4528 passed, 1 skipped; both trees untouched; cli/.the-loop/ not created
T3	final conftest.py over the unfixed tests and runtime code	74 passed, 7 errors: exactly the four writers, each naming its path
T4	tests/test_state_isolation.py	11 passed, from both directories
T5	the four fixed modules + test_channels_dm*.py	green (inside T1/T2; 120 passed when run alone)
T6	make lint format-check typecheck validate	green: ruff, markdownlint (0 errors), pyright (0 errors), configs valid
T7	full suite from cli/, without then with the hook, back to back	185.05s without, 180.81s with: no measurable cost
T15	sitecustomize tracer over a root run	three of the four writers (it did not wrap mkdir), all in the test process, none in a child
AC3	git ls-files .the-loop/portable	empty

How the writers were found (T15)

A sitecustomize.py on PYTHONPATH wrapped builtins.open, io.open, os.open, os.replace and tempfile.mkstemp. For every write under either .the-loop/, it logged the PYTEST_CURRENT_TEST, the pid and argv, and the stack. Because it is loaded by PYTHONPATH, every child that inherits the test's environment loads it too. Run on main from the root, every logged write was in the pytest process itself, grouped as:

1 test=tests/test_api_health_integration.py::test_a_hosted_loop_that_ends_on_its_own_drops_its_lock_and_says_so (call) os.open /home/user/the-loop/.the-loop/poll.pid
2 test=tests/test_graph_drive_integration.py::test_the_spawning_event_reaches_the_graph (call) mkstemp .the-loop/portable
2 test=tests/test_graph_drive_integration.py::test_the_spawning_event_reaches_the_graph (call) os.open /home/user/the-loop/.the-loop/portable/TMP
1 test=tests/test_graph_drive_integration.py::test_the_spawning_event_reaches_the_graph (call) replace .the-loop/portable/github-octo-repo-15.json
1 test=tests/test_graph_drive_integration.py::test_the_spawning_event_reaches_the_graph (call) replace .the-loop/portable/index.json
2 test=tests/test_sessions_restart_integration.py::…[failed] (call) open .the-loop/logs/events.jsonl
2 test=tests/test_sessions_restart_integration.py::…[fresh] (call) open .the-loop/logs/events.jsonl
2 test=tests/test_sessions_restart_integration.py::…[stale] (call) open .the-loop/logs/events.jsonl
2 test=tests/test_sessions_restart_integration.py::…[unverified] (call) open .the-loop/logs/events.jsonl

The tracked record's stack, trimmed to the-loop's frames:

File "cli/tests/test_graph_drive_integration.py", line 284, in test_the_spawning_event_reaches_the_graph
  dispatcher.handle(routed)
File "cli/the_loop/webhook/dispatcher.py", line 1152, in handle
  self._apply_control(control.command, routed)
File "cli/the_loop/webhook/dispatcher.py", line 2261, in record
  self.control_store.record(
File "cli/the_loop/control.py", line 636, in record
  self.store.write_section(item, CONTROL, record.to_dict())
File "cli/the_loop/workitem.py", line 398, in _write_json
  fd, tmp_name = tempfile.mkstemp(dir=str(self.root), suffix=".tmp")

Every write is in a test's (call) phase. None is at session end, which settles the ticket's shutdown hypothesis. The tracer did not wrap os.mkdir, so it missed the fourth writer (test_doctor_slack, which creates local/). The guard's first run (T3) found it.

T3: negative control

The final cli/tests/conftest.py was copied into a clean worktree of main (a754cea), with the tests and runtime code unfixed, and the four modules were run from its root:

$ python -m pytest -q cli/tests/test_api_health_integration.py cli/tests/test_doctor_slack.py \
    cli/tests/test_graph_drive_integration.py cli/tests/test_sessions_restart_integration.py
wrote into a protected `.the-loop/` state tree — give the code under test a state root under tmp_path (issue-422):
os.mkdir <worktree>/.the-loop
open <worktree>/.the-loop/poll.pid
os.remove <worktree>/.the-loop/poll.pid
wrote into a protected `.the-loop/` state tree — …
os.mkdir .the-loop/local
open <worktree>/.the-loop/local/tmpu_ki4v0g.tmp
os.rename .the-loop/local/slack-directory.json
wrote into a protected `.the-loop/` state tree — …
os.mkdir .the-loop/portable
os.rename .the-loop/portable/github-octo-repo-15.json
os.rename .the-loop/portable/index.json
wrote into a protected `.the-loop/` state tree — …   (×4, one per restart case)
os.mkdir .the-loop/logs
open .the-loop/logs/events.jsonl
ERROR cli/tests/test_api_health_integration.py::test_a_hosted_loop_that_ends_on_its_own_drops_its_lock_and_says_so
ERROR cli/tests/test_doctor_slack.py::test_doctor_slack_prints_all_three_sections_and_exits_zero_when_clean
ERROR cli/tests/test_graph_drive_integration.py::test_the_spawning_event_reaches_the_graph
ERROR cli/tests/test_sessions_restart_integration.py::TestSecretsAreNeverPrinted::test_no_declared_value_reaches_any_output_on_any_path[fresh]
ERROR cli/tests/test_sessions_restart_integration.py::TestSecretsAreNeverPrinted::test_no_declared_value_reaches_any_output_on_any_path[stale]
ERROR cli/tests/test_sessions_restart_integration.py::TestSecretsAreNeverPrinted::test_no_declared_value_reaches_any_output_on_any_path[unverified]
ERROR cli/tests/test_sessions_restart_integration.py::TestSecretsAreNeverPrinted::test_no_declared_value_reaches_any_output_on_any_path[failed]
74 passed, 7 errors in 3.80s

The same four tests, the same seven cases, went red in the full-suite run from the repository root on this checkout before task A2 (4517 passed, 1 skipped, 7 errors).

T1 and T2: both working directories, after the fix

At 06e06e6:

## T1 root
$ git status --porcelain --ignored .the-loop cli/.the-loop
$ uv run --project cli python -m pytest -q cli
4528 passed, 1 skipped in 200.83s (0:03:20)
$ git status --porcelain --ignored .the-loop cli/.the-loop
$ git status --porcelain
## T2 cli/
$ cd cli && uv run python -m pytest -q
4528 passed, 1 skipped in 177.61s (0:02:57)
$ git status --porcelain --ignored .the-loop cli/.the-loop
$ git status --porcelain
$ git ls-files .the-loop/portable

Every status is empty, --ignored included, so neither the tracked records nor the ignored local/, logs/ and poll.pid were written. On main, the same root run reports the record modified (M) and .the-loop/local/ and .the-loop/logs/ as ignored (!!).

T4: the guard's own tests

$ cd cli && uv run python -m pytest -q tests/test_state_isolation.py
11 passed

Two of these were written red first, during self-review, and fixed in the hook:

E  AssertionError: assert [('os.symlink', '.the-loop')] == []
E  AssertionError: assert [('os.rmdir', … '.the-loop')] == []

T7: cost

The same machine, one run after the other, from cli/. The first run used a clean worktree of main, whose conftest.py has no hook. The second used this branch.

4517 passed, 1 skipped in 185.05s (0:03:05)   # main, no hook
4528 passed, 1 skipped in 180.81s (0:03:00)   # this branch, hook installed (+11 guard tests)

The difference is within run-to-run noise. T1 and T2 above (200.83s, 177.61s) show the same spread between directories.

T12: the restored abuse case

test_no_declared_value_reaches_any_output_on_any_path now asserts (tmp_path / "events.jsonl").read_text() is non-empty before searching it. All four cases pass, so the search covers a real log and finds no declared value in it.