Orchestrata Ultimate › Plans › briefs › qwen › Q8-continuation

Q8 continuation — the leaking live_run_tree fixture, then finish

Source: plans/briefs/qwen/Q8-continuation.md at 8a2cb42ad · last changed 2026-10-12 · first committed 2026-10-12 · 73 lines.

Prose pending.

Plan text

Q8 continuation — the leaking live_run_tree fixture, then finish

Same worktree, same branch, same rules as plans/briefs/qwen/Q8-hook-process-matching.md. Do not commit. Every scripts/run_tests.py run goes through bash .run_exclusive.sh .... Every measured step goes through bash .hostlock.sh ..., if your worktree has it. A guard denial is reported, never worked around.

1. The defect (orchestrator, measured)

tests/claude/test_test_run_context.py, fixture live_run_tree (lines 74-111 on master):

• It starts a parent: python -c ... scripts/run_tests.py.
• The parent spawns a child: python -c 'import time; time.sleep(60)' -m pytest.
• Teardown kills only the parent: parent.kill(); parent.wait(...).
• The child is re-parented to pid 1 and lives up to 60s after the test. Every test that uses the fixture leaks one such process.

Evidence:

• 2026-10-11 21:20: 14 orphans were alive after a run of tests/claude. Each had ppid 1 in /proc/<pid>/stat and cwd wt-COVERAGE-GATE.
• 21:31: four more, e.g. pid 1158826: /home/mohamed/workspace/wt-COVERAGE-GATE/.venv/bin/python -c import time; time.sleep(60) -m pytest, ppid 1.

Anything that scans /proc sees these as pytest processes:

• the status hook you are fixing (Part 2);
• your host-safety proof (Verification 2);
• any concurrent test that enumerates pytest processes.

2. Fix, red first

1. Factor, do not replace.
   • Move the fixture's body into a module-level context manager, _live_run_tree() (contextlib.contextmanager). It yields the same object the fixture yields today, plus a way to read the child's pid; add that only if it is not already exposed.
   • The fixture keeps its name, scope and yield value, and becomes with _live_run_tree() as tree: yield tree.
   • No existing test changes, and no assertion changes.
2. Red test first. Add a test that uses the context manager directly:

   def test_live_run_tree_leaves_no_child_behind() -> None:
       with _live_run_tree() as tree:
           child = <the child pid>
           assert Path(f"/proc/{child}").exists(), "precondition: the child is running inside the block"
       deadline = time.monotonic() + 5
       while Path(f"/proc/{child}").exists() and time.monotonic() < deadline:
           time.sleep(0.1)
       assert not Path(f"/proc/{child}").exists(), f"child {child} outlived the fixture"
   

   • Run the module through the wrapper before the fix. Quote the red output: the child outlives the block.
   • A zombie (State: Z in /proc/<child>/status) still has a /proc entry. If the red result depends on that, say so and quote /proc/<child>/status.
3. The fix, inside _live_run_tree():
   • start the parent with start_new_session=True;
   • at teardown, os.killpg(parent.pid, signal.SIGKILL) (tolerate ProcessLookupError), then parent.wait(timeout=...);
   • then poll /proc/<child_pid> for up to 5s and raise AssertionError if it remains. A leak becomes a red result instead of a silent one.
4. Same treatment for any other fixture in the files you touch that spawns grandchildren. Check foreign_run. Report each fixture you checked, and what you found.
5. Run the module again. It must be green; quote the totals.

3. Verification additions

• Do Verification 2 (host-safety proof) after this fix.
• Also report that no -m pytest orphan with ppid 1 and your worktree as cwd remains after the q8 run. Run this and quote its output (empty is the pass):

  for p in $(pgrep -f -- '-m pytest'); do [ "$(awk '{print $4}' /proc/$p/stat 2>/dev/null)" = 1 ] && readlink /proc/$p/cwd; done | grep -c wt-HOOK-MATCH
  

4. Finish

Complete your brief's done-when and write plans/briefs/qwen/Q8-report.md. Include this item with its red/green evidence and the orphan check output. Do not commit.

Related

• Plans
• Orchestrata Ultimate — project index
• Workflow
• Tasks