Repository navigation
flake: rpc-host-gc 'removes an uncleanly killed endpoint' finds nothing to remove (CI, unrelated PR) #4119
Description
Summary
packages/coding-agent/test/rpc-host-gc.test.ts: the case gc against real supervised hosts > removes an uncleanly killed endpoint, its socket and its siblings, and keeps the live one beside it failed once in CI on a PR that touches no RPC host or gc code.
Evidence
- Run: https://github.com/code-yeongyu/senpi/actions/runs/39971837908 (job
Test (coding-agent 3/3)), on fix(models): a models.json entry keeps the built-in price; an unpriced model costs unknown, not $0 #3847. - Assertion:
expected { removed: [], …(1) } to deeply equal { removed: [ { …(3) } ], …(1) }. gc ran, but did not remove the uncleanly killed endpoint. - fix(models): a models.json entry keeps the built-in price; an unpriced model costs unknown, not $0 #3847's diff has no file under
src/modes/rpc/and no host or gc code. Its only RPC-adjacent file ispackages/coding-agent/docs/rpc.md. - In the 8 most recent failed
CIruns on other branches, this case did not fail.
Expected
The case passes deterministically. gc removes the killed endpoint's record, socket and siblings, and keeps the live host.
Hypothesis (unverified)
The test drives real supervised host processes. If gc runs before the killed host's death is observable (its pid still answers, or the lock or socket is not released yet), gc correctly treats the endpoint as live and removes nothing. That would make the case timing-dependent on a loaded runner, like #2876 (a different case in the same file).
Related
- flake: rpc-host-gc 'recreated endpoint.json while gc held the lock' times out waiting for endpoint.json on a loaded host #2876 (load flake in the same file, different case).
Activity
code-yeongyu commentedon Oct 24, 2026
Sibling in the same family flaked on CI for an unrelated PR (senpi#4109, which touches only the subscription-limit store):
test/rpc-host-gc.test.ts > gc against real supervised hosts > removes an endpoint whose supervisor alone was SIGKILLed, after its host child's watchdog cleaned up, expected removed:[endpoint], got kept:[endpoint]. Run: https://github.com/code-yeongyu/senpi/actions/runs/40051880459/job/120746130698fix(rpc): treat an unreaped leader and claim owner as exited too (#4119…
code-yeongyu commentedon Oct 25, 2026
Fixed by #4160 (merge commit
f8eeec6a03).Root cause. gc read a SIGKILLed supervisor or host child as still running while it was exited but not yet reaped (state Z):
kill(pid, 0)succeeds for a zombie andps -o lstart=still prints its start time, so the generation pidfile matched as live and gc kept the dead endpoint aslive_generation. On a loaded runner the reap could land after gc read the pidfile, which is why this flaked rather than failed every time.Fix. The identity probe reads state and start time from one
pssnapshot and treats state Z as absent; pidfile matching, the host-child check, session-path claim ownership and process-group ownership all treat an unreaped process as exited. Process-group cleanup keeps its previous verdict for a zombie leader (sweep).Evidence.
host.pid/host-child.pidand as a claim owner: before the fix gc returnedremoved: []withlive_generation/live_claim, the diff in this issue; reverting only the source change makes every new case fail again.The separate flake #2876 (recreated
endpoint.jsoncase) is unrelated: an A/B under identical load showed the same rate with and without this change.fix(rpc): gc no longer counts an exited, unreaped host as a live gene…