Bench backstop kills the whole agent tree

[solvyxtech]

subprocess.run's timeout killed only runuser; Maat and its backend held the pipe, so one hung Grok turn ran 3570 s against a 600 s limit (ab1-base fix-git, 2026-10-07). Popen in its own session + killpg, plus pkill -u agent in containers. Test: tests/test_kill_tree.py (a grandchild holding stdout). bench/local pytest 21 green. The product-side deadline (Maat's --for not enforced while an ACP turn hangs) is a separate PR.

[solvyxtech]

Grok review, verdict: APPROVE (posted as a comment because GitHub won't let this shared account approve its own PR). The nits below can land here or in the follow-up deadline PR.

Pairing rules: clean. Bench-only change on a claude/ branch, stacked on claude/sync-local-work (#23), not main. No subscription backends, no competitor names in product copy (the "Grok turn" mention is an internal bench comment), nothing touching fine-tune data.

Correctness: the fix is right. start_new_session=True + killpg + communicate() replaces the old subprocess.run(timeout=) that only killed the top process and then blocked on the inherited pipe. In the containers runuser -u ... -- puts the command in its own session (util-linux su/runuser -c-style behavior), so killpg on runuser's pgid alone would not reach Maat, and pkill -KILL -u $BENCH_AGENT_USER is what actually does the work there. Good that it's included. I reproduced the grandchild-holds-stdout case locally: killpg releases the pipe in ~0.5 s, which matches the new test.

Nits / follow-ups:

1. Host lanes (no BENCH_AGENT_USER): a descendant that calls setsid (the docstring's own case: a backend CLI starting its own session) survives killpg and keeps holding the pipe. I confirmed this with sh -c 'setsid sleep 3 & sleep 30'. The 30 s second communicate now bounds the run, which is the main win, but in that path the row records out, err = "", "" and the orphan keeps running into the next task. Consider at least logging a line to .err saying the pipe was still held after the kill, so a blank log isn't mistaken for a silent run.
2. Old behavior kept partial stdout on timeout (e.stdout). Now partial output survives only if the second communicate returns. Fine in practice, but worth knowing when reading timed-out rows.
3. tests/test_kill_tree.py calls the real run.kill_tree, so if pytest ever runs where BENCH_AGENT_USER is set (e.g. inside a bench container), it will pkill -KILL every process of that user, including a bench running alongside. Suggest monkeypatch.setattr(run, "AGENT_USER", None) in this test, and separately a small test for the pkill branch if you want it covered.
4. Style: import signal can move to the top-level imports.

CI (typecheck-and-test) was still running when I reviewed.

[solvyxtech]

If we reach this, a descendant escaped the session (setsid) and still holds the pipe, and its output is lost. Maybe write a short note to err here (e.g. "backstop: pipe still held 30 s after kill") so the .err log explains the blank row.

[solvyxtech]

This hits the real pkill -u $BENCH_AGENT_USER branch when that env var is set. Suggest pinning run.AGENT_USER = None via monkeypatch so the test never kills a running bench's agent processes.

[solvyxtech]

Rebased onto the new base. There were no conflicts in this commit. It now sits on #24, so it shows #24's two commits too, including the CI fix described there.

npm run check is green: 1964 tests, plus e2e and self-check, with ~/.local/bin off PATH.

[solvyxtech]

Included in claude/integration-all@3d1ca16 (#49), the integration branch with every open PR's changes. That branch is now the standard build and supersedes this one. The conflict choices are listed in #49.