fix(cli): kill spawned claude binary when launcher exits (orphan / 'two cursors' bug)

[johnmarktaylor91]

Symptom

When happy is installed against a binary claude (Homebrew or the native installer at ~/.local/share/claude/versions/<v>/), users intermittently report that typing produces characters at two alternating cursor positions in the same pane, making input impossible. Often described as "two cursors" or "tmux is broken." It isn't tmux.

Root cause

runClaudeCli in packages/happy-cli/scripts/claude_version_utils.cjs has two branches:

• .js/.cjs claude: import(importUrl) -- runs in-process, no orphan possible.
• Binary claude: spawn(cliPath, args, { stdio: 'inherit' }) with only a child.on('exit', ...) handler. No signal forwarders, no parent-death detection.

When the launcher process dies for any reason -- terminal SIGHUP on pane close, kill -TERM, an OOM, an uncaught exception in the launcher's fetch interceptor, or the parent (happy) crashing without sending signals down the chain -- the spawned claude binary keeps running. Linux reparents it to PID 1, but it stays in the TTY's foreground process group (STAT=Sl+).

Next time happy launches a session in the same pane, a second claude binds to the same /dev/pts/N. Both are in the foreground PG, so the kernel race-distributes each keystroke between them. That's the "two cursors" symptom.

Reproduction on a machine with a binary claude install:

ps -eo pid,ppid,stat,tty,cmd | grep -E 'claude/versions' | grep -v grep
# look for: PPID=1, STAT=Sl+, multiple instances on same pts/N

Fix

In the binary branch of runClaudeCli:

1. Forward SIGINT/SIGTERM/SIGHUP/SIGQUIT from launcher to the spawned child.
2. On launcher 'exit', send SIGTERM to the child (covers natural exit).
3. 1Hz ppidWatch that fires when the launcher's process.ppid === 1 (covers ungraceful parent death where no signal is propagated). Interval is unref()'d so it doesn't keep the event loop alive after normal child exit.

21 lines added, no behavior change for the existing happy path.

Verified locally

Synthetic launcher mirroring runClaudeCli's spawn pattern, with a long-lived /bin/sleep as the fake child:

Test	Scenario	Result
1	UNPATCHED control: kill -TERM launcher	child alive, PPID=1 (bug reproduced)
2	PATCHED + kill -TERM launcher	child dies
3	PATCHED + kill -HUP launcher (terminal-close case)	child dies
4	PATCHED + kill -KILL launcher's parent shell	child dies within 1s via ppid watch

Notes

• Doesn't change the .js/.cjs import-path branch, since orphans are impossible there.
• Could alternatively use prctl(PR_SET_PDEATHSIG) via FFI for instant detection, but that's Linux-only and adds a native dep. The 1Hz polling is portable and adds negligible CPU.
• I'm happy to add a test if you want one -- the synthetic harness above is straightforward to port into the test suite.