daily-fix: spawn stop --kill must TERM inner claude pid
Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.
No comments yet.
Workflow gap
On the daemon-untracked zombie path, spawn_session.py stop --kill SIGTERMs the wrapper
and exits 2 with "WARNING: inner claude pid still alive — the wrapper's SIGTERM
cleanup may have failed; verify/kill manually". Observed live 2026-08-05T16:10–16:11Z
killing the ~24 h zombie holding #2061 (wrapper node pid 2943628, ELAPSED 1-00:27:51;
plain stop had already failed with "Raw daemon reply: {'success': False}"). The inner
pid happened to be gone by the manual check, but the tool's contract leaves a live inner
claude process to a human race — on an autonomous fleet the "verify/kill manually" step
has no owner.
verified-at-filing: the exit-2 warning text + the failed plain-stop reply are probed
tool_result rows (session a4155180 rows 92–104). grep -n 'still alive' scripts/spawn_session.py | head -3
run at compose time — the warning path is live; no bounded-wait/TERM follow-up after the
wrapper SIGTERM.
Proposed change
In scripts/spawn_session.py's daemon-untracked --kill path: after SIGTERMing the
wrapper, bounded-wait (~10 s) for the inner claude pid; if still alive, TERM it directly,
re-verify, and only then exit — exit 2 with the manual instruction remains the LAST
resort when the inner pid survives TERM (never auto-escalate to KILL; a KILL on a live
session is the destructive case a human should confirm). Also consider a watcher arm for
daemon-untracked-but-live wrappers on blocked tasks (this zombie held #2061 blocked ~18 h
before anyone noticed).
Provenance
-
fingerprint: e4d5ff32fea0
-
workflow_fix_target: scripts/spawn_session.py
-
origin: /daily 2026-08-05 problem sweep — miner 5 P17 (probed rows).