app-server brokers never self-terminate → orphaned process/RAM leak (34 chains, 272 procs, ~2.2GB)

[cooperbi159]

Bug: app-server brokers never self-terminate → orphaned process/RAM leak

Repo: openai/codex-plugin-cc
Version: codex plugin 1.0.6
File: scripts/app-server-broker.mjs (+ context in scripts/lib/broker-lifecycle.mjs, scripts/session-lifecycle-hook.mjs)

Summary

The shared app-server broker has no idle timeout. It exits only on an explicit
broker/shutdown RPC, SIGTERM, or SIGINT. Its only automatic reaper is the
SessionEnd hook, which is fragile. When that hook fails to run, the detached
broker (child.unref()) leaks forever, holding a Codex app-server child plus
its MCP/worker subprocesses.

Observed in the wild: 34 orphaned broker chains, 272 processes, ~2.2 GB RSS,
0 clients connected. 16 chains pointed at already-deleted git worktrees.

Root cause

Two gaps compound:

1. Broker never self-exits. In app-server-broker.mjs, socket.on("close")
   just does sockets.delete(socket) + clearSocketOwnership(socket). There is
   no check for "no clients left" and no idle timer. A broker with zero clients
   lives indefinitely.

2. The reaper is best-effort and misses common cases. SessionEnd in
   session-lifecycle-hook.mjs is the intended reaper, keyed per-cwd via
   loadBrokerSession(cwd). It fails to reap when:

   • Claude is hard-killed / crashes → hook never runs.
   • The 5s hook timeout (hooks.json) is exceeded on a busy machine → hook killed.
   • The worktree is deleted before SessionEnd → loadBrokerSession(cwd)
     resolves to gone state, returns null, no shutdown is sent → orphan.
     (This matches the 16 deleted-worktree chains.)

Because brokers are spawned detached and keyed to a cwd, any hook miss orphans a
broker that nothing will ever clean up.

Fix

Give the broker an idle self-exit. It is safe: the companion opens one socket
per CodexAppServerClient.connect() and closes it when the job ends
(BrokerCodexAppServerClient.initialize/close in lib/app-server.mjs), so a
broker at zero sockets with no active request/stream is genuinely idle. Killing
an idle broker only costs a cold respawn on the next job — ensureBrokerSession
already checks endpoint readiness and respawns if dead. This makes the broker
self-healing regardless of whether any hook fires, covering hard-kills and
deleted worktrees alike.

Threshold defaults to 15 min, overridable via CODEX_COMPANION_BROKER_IDLE_MS.

Patch (scripts/app-server-broker.mjs)

Add idle state after const sockets = new Set();:

  const IDLE_TIMEOUT_MS = Number(process.env.CODEX_COMPANION_BROKER_IDLE_MS ?? 15 * 60 * 1000);
  let idleTimer = null;
  let serverRef = null;

  function isIdle() {
    return sockets.size === 0 && !activeRequestSocket && !activeStreamSocket;
  }

  function clearIdleTimer() {
    if (idleTimer) {
      clearTimeout(idleTimer);
      idleTimer = null;
    }
  }

  function armIdleTimer() {
    clearIdleTimer();
    if (!(IDLE_TIMEOUT_MS > 0) || !serverRef || !isIdle()) {
      return;
    }
    idleTimer = setTimeout(async () => {
      if (!isIdle()) {
        return;
      }
      await shutdown(serverRef);
      process.exit(0);
    }, IDLE_TIMEOUT_MS);
    if (typeof idleTimer.unref === "function") {
      idleTimer.unref();
    }
  }

Cancel the timer on connect (in net.createServer callback, after sockets.add(socket)):

    clearIdleTimer();

Re-arm on disconnect (in both socket.on("close") and socket.on("error"),
after clearSocketOwnership(socket)):

      armIdleTimer();

Arm at startup and hold the server ref (replace server.listen(listenTarget.path);):

  serverRef = server;
  server.listen(listenTarget.path, () => {
    armIdleTimer();
  });

Optional hardening (not required if idle-exit lands)

• Make the SessionEnd hook timeout longer than 5s, or move broker teardown off
  the hook's critical path, so slow machines still reap on clean exit.
• On SessionStart, sweep os.tmpdir() for stale cxc-*/broker.pid whose pid is
  dead (leftover session dirs). Redundant once idle-exit ships, but cleans state.

[HoneyTyagii]

Can I work on this?

[ghbaud]

Confirming the issue.