openclaw - 💡(How to fix) Fix camofox-browser plugin: autoStart child process killed by SIGTERM ~18s after launch [1 participants]

Bug: camofox-browser autoStart child process killed by SIGTERM ~18s after launch

OpenClaw version: 2026.4.15 OS: macOS Darwin 25.4.0 (arm64) Plugin: camofox-browser v1.5.2 (global, enabled) Node: v25.8.1

Summary

When the camofox-browser plugin autoStarts its server via launcher.js, the child process receives SIGTERM approximately 15–20 seconds after launch (mean 18.5s across 17 samples). The server never survives past autoStart.

However, running node server.js directly (outside the Gateway plugin lifecycle) results in a stable, long-running process with no SIGTERM.

Steps to Reproduce

1. Enable camofox-browser plugin with autoStart: true (default)
2. Start or restart OpenClaw Gateway
3. Observe camofox-server.log: server started → browser pre-warmed → ~18s → shutting down, signal:"SIGTERM"

Evidence

17 consecutive SIGTERM samples (start → kill interval):

• Mean: 18.5s, Min: 0.5s, Max: 33.3s
• Most samples cluster at 15–20s (e.g. PID 44700: 17.6s, 44814: 18.7s, 45203: 17.5s, 45960: 19.8s, 46782: 16.7s)

Log sequence (repeated every cycle):

{"ts":"...","level":"info","msg":"server started","port":9377,"pid":XXXXX}
{"ts":"...","level":"info","msg":"browser pre-warmed","ms":310}
... ~18 seconds gap ...
{"ts":"...","level":"info","msg":"shutting down","signal":"SIGTERM"}

Control Experiments

Ruled Out

• ❌ Server internal idle shutdown (only closes browser, not server; idle timeout is 300s)
• ❌ plugin.ts主动 kill (no kill/SIGTERM logic in plugin code)
• ❌ gateway-watchdog (runs every 600s, not 18s)
• ❌ launchd (no agent targeting port 9377 or server.js)
• ❌ macOS kernel (direct start is stable)

Root Cause Hypothesis

The Gateway plugin lifecycle/supervision layer (in compiled dist chunks) appears to kill child processes spawned by plugins. Despite detached: true + proc.unref() in launcher.js, the stdio: ['ignore', 'pipe', 'pipe'] keeps pipe connections to the parent process. The SIGTERM arrives consistently ~18s after the autoStart async IIFE completes, suggesting a delayed cleanup or health-check-based kill in the Gateway core.

launcher.js (modified locally to add detached + file logging)

const proc = startProcess('node', args, {
  cwd: pluginDir,
  env: { ...env, CAMOFOX_PORT: String(port) },
  stdio: ['ignore', 'pipe', 'pipe'],
  detached: true,
});
proc.unref();

Suggested Fix

Either:

1. Gateway plugin lifecycle should respect detached: true + unref() and not kill child processes
2. Or provide a plugin API for long-running subprocess management that isn't subject to supervision cleanup