Repository navigation
mcpx(daemon): full-package test run is killed by SIGTERM — an in-process server's delayed self-signal terminates the test binary #5262
Description
mcpx(daemon): full-package test run is killed by SIGTERM — an in-process server's delayed self-signal terminates the test binary
Observed (verified this run, laptop, origin/main 1ed5d03)
go test ./internal/daemon/ in pkgs/mcpx fails reproducibly with the test
binary itself dying of SIGTERM:
- Full suite
go test ./...:FAIL internal/daemon 38.699s— output is just
signal: terminated, no failing test named. - Standalone
go test -count=1 ./internal/daemon/: same result at 16.6s,
31.5s, 36.0s and 68.3s across four runs (the variance is the tell: the
signal lands on whatever test happens to be running). - A clean isolated re-run (no other mcpx test suite active on the machine)
still fails the same way at 36.0s. - Individual tests pass alone, e.g.
go test -run TestAHiddenToolCannotBeCalled ./internal/daemon/→ PASS in 0.41s (in the-vfull-package run that same
test was the one running when the binary died).
The other red packages in the same full-suite run are already tracked
(internal/settings #5242/#2716, internal/conformance #2716,
internal/defaults #5211); this daemon failure is not covered by any open
issue (searched: "daemon terminated", "signal terminated", "mcpx daemon
test").
Likely mechanism (leads, not proven)
internal/daemon/server.go:1518 handleShutdown — and the restart handler
at :1530 — schedule a delayed self-signal in a goroutine:
go func() {
time.Sleep(defaults.DaemonRestartSettle)
p, _ := os.FindProcess(os.Getpid())
_ = p.Signal(syscall.SIGTERM)
}()In production the server is the daemon process, so this is its graceful
exit. In tests the server runs in-process inside the test binary: any test
that drives a shutdown/restart request (or leaves one of these goroutines
pending past its own end) SIGTERMs the entire test process
DaemonRestartSettle later, mid-way through an unrelated test. Tests in
this package also register os.Getpid() as a fake daemon PID
(paths_test.go:333,406,428; takeover_test.go:224), so stop/takeover
paths that signal PIDs read from info records are a second suspect route
to the same outcome.
Why it matters
- The daemon package's tests cannot be trusted as a gate: the package is
red on main for a reason unrelated to any assertion, and the failure
report names no test, so real regressions hide inside it. - The same in-process self-signal pattern is a hazard for any embedder
that runs the daemon server inside a longer-lived process.
Acceptance criteria
go test -count=1 ./internal/daemon/passes repeatedly (say 5
consecutive runs) on a machine with no other mcpx suite running, and
also when run as part ofgo test ./....- The shutdown/restart self-signal is suppressed or redirected when the
server runs in-process for tests (e.g. an injectable exit hook, like
the existingexecSelfseam at server.go:1527), without changing the
production daemon's exit behaviour. - A regression test proves that handling a shutdown request in-process
does not deliver SIGTERM to the host process.
Activity
dezren39 commentedon Nov 4, 2026
🤖 Agent note · state: filed
go test ./internal/daemon/(pkgs/mcpx) kills its own test binary with SIGTERM on clean main. Verified on the laptop at origin/main 1ed5d03: full suite FAIL 38.7s, four standalone-count=1runs FAIL at 16.6/31.5/36.0/68.3s (incl. one clean isolated run), output onlysignal: terminated; individual tests pass alone (TestAHiddenToolCannotBeCalled 0.41s). Lead:internal/daemon/server.go:1518handleShutdown (and :1530 restart) schedule a delayedos.Getpid()SIGTERM afterdefaults.DaemonRestartSettle— correct for a real daemon process, fatal when the server runs in-process in the test binary.execSelf, server.go:1527) so in-process servers never SIGTERM their host; done whengo test -count=1 ./internal/daemon/passes 5 consecutive runs and as part ofgo test ./..., with a regression test that an in-process shutdown request delivers no SIGTERM to the host process.dezren39 commentedon Nov 4, 2026
agent-note:v1
task: 9108
kind: small_fix
state: picked_up
lane: code-h (direct execution)
next: reproduce on the laptop at current main, then route the shutdown/restart self-signal through an injectable seam in pkgs/mcpx/internal/daemon/server.go so an in-process server never SIGTERMs its host; regression test + 5x daemon suite runs.
dezren39 commentedon Nov 4, 2026
agent-note:v1
task: 9108
kind: small_fix
state: parked (CI pending — runner queue)
lane: code-h (direct execution)
done: fix implemented on branch queue/9108-5262-daemon-sigterm (commit 2859c03); PR #5318 open with "Fixes #5262". signalSelf seam routes both self-SIGTERMs (shutdown + restart fallback); production value is byte-for-byte the old behaviour. Regression test TestHandleShutdownSignalsSelfThroughSeam added; restart/handoff tests capture via the seam instead of signal.Notify.
evidence: go test -count=1 ./internal/daemon/ — 5 consecutive passes on the laptop (62.6/48.7/57.5/66.0/99.9s); go vet + gofmt clean. PR checks (mcpx-fix-carries-a-test, line-limit) still QUEUED after a 5-minute poll — same runner starvation as PR #5277's line-limit run.
next: on resume, read PR #5318 checks + Codex findings fresh; merge when green, then complete task 9108.
15 remaining items
dezren39 commentedon Nov 4, 2026
🤖 Agent note · parked
go test -count=1 ./internal/daemon/passed 5 consecutive times on the laptop. Note on ownership: PR fix(mcpx): route daemon self-SIGTERM through an injectable seam #5318 carries noheld,user/owned,do-not-merge, ordo-not-auto-editlabel and was opened by lane code-h for task 9108, so author name alone is not treated as a Drew-owned hard stop.dezren39 commentedon Nov 4, 2026
🤖 Agent note · picked up
fix(mcpx): route daemon self-SIGTERM through an injectable seam (#5318)
dezren39 commentedon Nov 4, 2026
🤖 Agent note · done