`npm test` leaves a detached broker and a fake app-server behind for every test workspace

What happens

Every codex-companion invocation that reaches the app-server calls CodexAppServerClient.connect(cwd), which starts a detached, unref'd app-server-broker.mjs serve for that workspace unless CODEX_COMPANION_APP_SERVER_ENDPOINT is set or disableBroker is passed (plugins/codex/scripts/lib/app-server.mjs, connect; broker-lifecycle.mjs, ensureBrokerSession). The broker keeps its codex app-server child alive, has no idle timeout, and only stops on broker/shutdown, which the SessionEnd hook sends.

The test suite creates a fresh temp workspace per test and never sends broker/shutdown, so each full run leaves roughly a hundred broker + fake-app-server pairs parented to launchd. On a developer machine they accumulate silently: after eight full runs on macOS I counted 1,108 fake codex app-server fixtures and 555 brokers, 36 GB resident, all idle at 0 % CPU. Nothing in CI notices because the runner is thrown away.

Observed on main at db52e28, node 22.17.0, macOS.

Reproduction

npm test >/dev/null 2>&1
ps -axo pid,ppid,command | grep -c "codex-plugin-test-.*/codex app-server"
# ~100 per run, ppid 1, still there an hour later

Possible fix

• A t.after in tests/helpers.mjs or the fixture that reads the broker session file for each temp workspace (loadBrokerSession(cwd)) and sends broker/shutdown, or kills the recorded pid; or
• a suite-wide env such as CODEX_COMPANION_APP_SERVER_ENDPOINT=unix:/nonexistent.sock so withAppServer falls back to the spawned client, for the tests that do not exercise the broker; or
• an idle timeout in the broker (say 30 minutes without a client), which would also help real sessions whose SessionEnd hook never runs (crashed terminal, killed IDE).

Start with tests/helpers.mjs and the fixture paths around loadBrokerSession(cwd), then run npm test and the supplied process check to reproduce the orphaned broker/app-server pairs. Trace the existing broker/shutdown path in plugins/codex/scripts/lib/app-server.mjs and broker-lifecycle.mjs. Done means the test run leaves no per-workspace broker or fake app-server processes.