fix(ctl): make the in-box ctl-daemon launch idempotent (stop the pile-up)

[madarco]

Problem

A live cloud box accumulated many agentbox-ctl daemon processes (a Hetzner box had 9), only the oldest actually serving the box-relay port :8788; the rest were orphaned idle node processes.

launchCloudCtlDaemon spawns the daemon unconditionally on every create / start / unpause / resume / recover (via reEnsureCloudBox). The daemon has no singleton guard — when :8788 is already bound it catches EADDRINUSE and keeps running — so each relaunch leaks another idle daemon. agentbox recover, which targets already-running boxes, hit this every single time. (It is not agentbox attach — attach never launches a daemon.)

Fix

• Cloud (ctl-launch.ts): guard the spawn with a liveness probe — if a healthy daemon already serves the box relay's /healthz (port parsed from relayUrl, default 8788), skip the launch. It already has the right env (stable across a host-only reconnect); only a real sandbox restart kills it (probe fails → relaunch). Probe via node (guaranteed present; curl may not be).
• Docker parity (ctl.ts): same idea with pgrep -f 'agentbox-ctl daemon$' (end-anchored so it matches the daemon, not the sh -c wrapper). Docker doesn't pile up today because the daemon dies with the container, but the new docker reconnect runs startBox on a possibly-running container and would otherwise double-launch.
• Corrected the stale "Idempotent — the launch* helpers no-op" comment in reEnsureCloudBox to match reality.

Verification

• Unit: assert the cloud script contains the /healthz skip-guard before the spawn, and the probe port is parsed from relayUrl.
• Live (Hetzner box): ran recover --no-attach twice — daemon count stayed flat, :8788 still healthy on the original pid.
• pnpm typecheck && lint && test green.

Note: this prevents new pile-up; existing orphans in a long-lived box clear on its next real restart. Out of scope: recover --adopt against a box with a foreign daemon (different token) needs force-replace — tracked as a follow-up.

https://claude.ai/code/session_01Ja5HgEjwyER5BhhFCpPUup

---

Note

Medium Risk
Changes in-box daemon startup on every start/resume/recover/reconnect path; mis-probes could skip needed relaunches, though fresh sandbox boots still relaunch when health checks fail.

Overview
Stops duplicate agentbox-ctl daemon processes when reEnsureCloudBox runs on an already-live box (notably recover / reconnect), where each unconditional spawn could leave orphaned Node processes after EADDRINUSE on the box-relay port.

Cloud (ctl-launch.ts): before nohup agentbox-ctl daemon, the launch script probes http://127.0.0.1:<port>/healthz via node (port from relayUrl, default 8788). A 200 skips the spawn; only a dead/missing daemon proceeds to mkdir, box.env, and launch.

Docker (ctl.ts): same idea with pgrep -f 'agentbox-ctl daemon$' so reconnect / start on a running container does not double-launch.

Comments in reEnsureCloudBox now state that wake paths include reconnect() and that ctl/dockerd/VNC launches are idempotent when already healthy.

^{Reviewed by Cursor Bugbot for commit 56fe054. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
agentbox-web	Skipped		Jun 29, 2026 9:29pm