Second, deeper pass over tools/gateway/hermes_cli plus first pass over
the trees wave 1 missed (acp, acp_adapter, skills, computer_use, docker,
dashboard, conformance, monitoring, secret_sources, hermes_state,
providers). Same rubric as wave 1 (AGENTS.md test policy); security,
alternation/caching invariants, issue-number regressions, and E2E kept.
Real test-quality fixes found and rooted out along the way:
- tests/tools/test_command_guards.py made real auxiliary-LLM HTTPS calls
(DEFAULT_CONFIG smart-approval leaked in) — pinned approval
mode=manual via autouse fixture: 17.4s → 0.4s.
- test_model_switch_custom_providers.py / test_user_providers_model_switch.py
silently probed live provider catalogs (~2s/test) — stubbed
cached_provider_model_ids/provider_model_ids/fetch_api_models.
- test_telegram_noise_filter.py: 15-platform copy-paste matrix over
shared gateway.run logic → 3 representative platforms (55s → 3.9s).
- test_gateway_shutdown.py: stop()'s 5s interrupt-deadline loop spun on
MagicMock agents — interrupt.side_effect now clears _running_agents
(22s → 1.0s).
- test_gateway_inactivity_timeout.py poll-harness timings shrunk 3-5x
(24s → 1.1s); test_mcp_stability.py backoff/SIGTERM-grace sleeps
patched (15.4s → 2.5s); test_async_delegation.py negative-drain wait
5s → 0.5s.
- test_telegram_init_deadline.py: loop-block margin restored to 1.0s
with rationale comment — the watchdog-dump assertion needs the loop
blocked well past deadline+grace under parallel load (flaked once in
the 40-worker verification run at a 0.2s margin).
Verification: full hermetic suite via scripts/run_tests.sh —
2,438 files, 21,718 tests passed, 0 failed, 293.9s wall.
Suite totals vs original baseline: 46,820 → 19,757 test functions
(−57.8%), wall 583.5s → 293.9s (−50%), subprocess CPU 13,564s → 11,623s.
The salvaged test spawned a listener subprocess that printed its port
immediately after bind() but BEFORE listen(), so under CI's loaded 8-worker
box the parent connected before the socket was listening -> ConnectionRefused
(flaked on test slice 2/6). Reorder the child to listen() then print the port,
and make the client connect with a short bounded retry to absorb scheduler
jitter. 15/15 green locally including direct hammering.
Follow-up to the salvaged #43846 commits: the WhatsApp adapter moved from
gateway/platforms/whatsapp.py to plugins/platforms/whatsapp/adapter.py since the
PR was authored. The cherry-pick brought _listener_pids_on_port's `re.finditer`
ss-fallback and the new test's import, but the new module location doesn't import
`re` (latent NameError on the lsof-absent fallback path) and the test imported the
old module path. Add `import re` to the adapter and repoint the test import.
This is the bug that was actually closing Firefox. `_kill_port_process`, run on
every bridge (re)start to free the port, used `lsof -ti :PORT` / `fuser PORT/tcp`
— both of which match a process whose socket merely *involves* that port number
in ANY state, including ESTABLISHED client connections. It then SIGTERMed every
match.
The bridge defaults to port 3000 — a ubiquitous local dev-server port. With a
browser tab open on localhost:3000, `lsof -ti :3000` returned Firefox's PID, so
each restart of the (crash-looping) WhatsApp bridge SIGTERMed Firefox, closing
the whole browser at irregular intervals with no crash and no coredump.
Proven live with the kernel `signal:signal_generate` tracepoint:
hermes-gateway(3396516) -> sig=15 (code=0/SI_USER) -> comm=firefox pid=3371585
captured immediately after a gateway start, while Firefox held a socket on the
bridge port. Demonstrated over-match: `lsof -ti :8080` returns the listener AND
the gateway's own client connection; `lsof -ti tcp:8080 -sTCP:LISTEN` returns
only the listener.
Fix: `_listener_pids_on_port` resolves only LISTEN-state sockets
(`lsof -ti tcp:PORT -sTCP:LISTEN`, with an `ss -ltnp` fallback) and
`_kill_port_process` signals just those. A client whose connection happens to
involve the port number is never touched — which is also more correct, since a
client never blocks the new bridge from binding. Windows already filtered
LISTENING; the broad `fuser -k` path is removed.
Adds TestKillPortProcess: real-socket tests proving a separate client process
is excluded from the listener lookup and survives port cleanup. 9 tests green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
`_kill_stale_bridge_by_pidfile` SIGTERMed the PID recorded in `bridge.pid`
after only a bare liveness check. Once the bridge exits and is reaped the
kernel recycles that PID onto an unrelated process; because the WhatsApp bridge
crash-loops ("Bridge process died (exit code 1)" repeating), this cleanup ran
on every restart and could SIGTERM a recycled PID that had landed on the user's
browser — closing Firefox at irregular intervals with no crash and no coredump
(a clean kill of a stranger).
Same PID-recycling class as the MCP reaper (7bd1f8a2d) and the process-registry
host-PID guard (e6a99cef2); this was the third, and most actively-fired, path.
Fix: `_write_bridge_pidfile` now also records the leader's kernel start time
(line 2). `_kill_stale_bridge_by_pidfile` re-validates identity via
`_bridge_pid_is_ours` before signalling — the (pid, start time) pair must match,
or for legacy single-line pidfiles the live cmdline must name `node` + this
session's unique path. A recycled PID (different start time / cmdline) is logged
and skipped, never signalled. Legacy pidfiles stay readable.
Adds TestWhatsappBridgePidfile: real-process tests proving a genuine bridge is
reaped while a recycled PID (start-time mismatch, or non-bridge cmdline) is
spared. 7 new + 108 gateway/registry tests green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Teknium
valentt