Oussama Gabouj

Oussama Gabouj

Co-founder & CTO @ Compresr (YC W26) | Efficient LLMs & context compression | EPFL MSc Data Science

Starting a Second Agent Killed the First One

Issue

Run the gateway’s agent launcher in Terminal 1, work happily. Open Terminal 2 and start a second session. Terminal 1’s session exits.

Reported as a gateway bug, and reasonably so — it only happened when running through our launcher. It was a launcher bug, and a self-inflicted one.

Root Cause

On startup, the launcher deleted all Claude IDE lock files.

The intent was sound: a crashed session leaves a lock file behind, and a stale lock blocks the next start. Clearing them made startup reliable — as long as exactly one session ever existed.

But the agent watches its own lock file and treats removal as a signal to shut down, which is the correct behaviour for it: a vanished lock means something else has claimed the session. So starting instance #2 deleted instance #1’s live lock, and instance #1 did precisely what it was designed to do.

The deletion could not distinguish “left over from a crash” from “held by a process running right now” — because it never asked. Both look identical on disk. The only difference is whether the owning process still exists, and that information was available the whole time.

Solution

Check whether the owning PID is still alive before removing a lock file. Only genuinely stale locks — those whose process is gone — get cleaned up.

Live sessions keep their locks, so multiple terminals can run agents simultaneously without interfering, and crash recovery still works because a dead process’s PID does not resolve.

💡 Takeaway

  • A lock file without liveness information is just a file. The PID is what makes it a lock; checking it is what makes cleanup safe.
  • “Clear all state on startup” scales to exactly one instance. It is a very common shape of fix that works perfectly until the second user, and then fails in a way that looks like someone else’s bug.
  • Interpret bug reports structurally. “The agent exits randomly” and “the agent exits when I open a second terminal” are the same defect; only the second one is debuggable, and getting there took asking what else was running.