fix: bound the lifetime of detached brokers and task workers - #652

Open

principalwater wants to merge 21 commits into

openai:mainfrom

principalwater:fix/reclaim-detached-process-trees

Commits on Aug 17, 2026

fix: bound the lifetime of detached brokers and task workers

principalwater
and
claude
committed
f7cd86a
fix: treat a zero limit as disabled when arming timers

principalwater
and
claude
committed
c24d41b
fix: clear the persisted broker session on idle shutdown

principalwater
and
claude
committed
44e427c
fix: probe a persisted broker session before reusing it

principalwater
and
claude
committed
b1371af
fix: tear down broker artifacts on idle shutdown, and recheck before deleting a record

principalwater
and
claude
committed
3ea94be
fix: tear down a stale broker's artifacts, but only once it is really gone

principalwater
and
claude
committed
98669f6
fix: keep the record of a broker that is slow rather than dead

principalwater
and
claude
committed
69a72b7
fix: do not wait for an exit that no transport will report

principalwater
and
claude
committed
f62edae
fix: do not reclaim a broker that is still running

principalwater
and
claude
committed
2ef1c57
fix: let a superseded broker clean up its own artifacts

principalwater
and
claude
committed
efe1cf2
fix: keep idle shutdown armed when a streaming client disconnects mid-turn

principalwater
and
claude
committed
957fbd1
fix: close the remaining hang and orphan paths around broker lifetime

principalwater
and
claude
committed
79790e2
fix: make the shutdown grace period actually bound the shutdown

principalwater
and
claude
committed
a397c88
fix: stop discarding a recorded shared runtime, and scope the fatal teardown

principalwater
and
claude
committed
1a6191f
fix: never abandon a turn that already finished

principalwater
and
claude
committed
d9e3065
fix: remember a completion only while its start awaits handoff

principalwater
and
claude
committed
da91429
fix: escalate to SIGKILL so the worker TTL is actually a ceiling

principalwater
and
claude
committed
3f8d7c5
fix: send a usable interrupt, and rearm idle shutdown when a request settles

principalwater
and
claude
committed
17649c2
fix: close both defect classes at every site, not one site at a time

principalwater
and
claude
committed
1b49b12
fix: match a racing completion by window rather than by thread id

principalwater
and
claude
committed
3180be7
fix: keep an expired job failed, and bound cleanup for a live transport

principalwater
and
claude
committed
efdfdeb

Navigation Menu

Uh oh!

fix: bound the lifetime of detached brokers and task workers - #652

fix: bound the lifetime of detached brokers and task workers#652
principalwater wants to merge 21 commits into
openai:mainfrom
principalwater:fix/reclaim-detached-process-trees

Commits on Aug 17, 2026

fix: bound the lifetime of detached brokers and task workers

fix: treat a zero limit as disabled when arming timers

fix: clear the persisted broker session on idle shutdown

fix: probe a persisted broker session before reusing it

fix: tear down broker artifacts on idle shutdown, and recheck before deleting a record

fix: tear down a stale broker's artifacts, but only once it is really gone

fix: keep the record of a broker that is slow rather than dead

fix: do not wait for an exit that no transport will report

fix: do not reclaim a broker that is still running

fix: let a superseded broker clean up its own artifacts

fix: keep idle shutdown armed when a streaming client disconnects mid-turn

fix: close the remaining hang and orphan paths around broker lifetime

fix: make the shutdown grace period actually bound the shutdown

fix: stop discarding a recorded shared runtime, and scope the fatal teardown

fix: never abandon a turn that already finished

fix: remember a completion only while its start awaits handoff

fix: escalate to SIGKILL so the worker TTL is actually a ceiling

fix: send a usable interrupt, and rearm idle shutdown when a request settles

fix: close both defect classes at every site, not one site at a time

fix: match a racing completion by window rather than by thread id

fix: keep an expired job failed, and bound cleanup for a live transport