macOS PTY master leak persists in 0.62.0 after completed shell commands (token-free repro) #29592

What happened?

Gemini CLI 0.62.0 on macOS retains one distinct /dev/ptmx master per completed ordinary shell command in a minimal token-free reproduction. This looks like a recurrence of #15945 after #25079. It exhausts the host PTY limit and prevents new terminal/login sessions (openpty: Device not configured).

On the affected host kern.tty.ptmx_max=999. Four long-lived Gemini processes held 85, 185, 222, and 79 distinct PTY device numbers respectively, with no direct child processes at the snapshot. These are distinct masters, not duplicate references to a few terminals. No MCP server or model call is needed for the reproduction below.

Minimal reproduction

Environment: Gemini CLI 0.62.0 (npm bundled distribution), Node v26.7.0, macOS 26.5.2 arm64. The bundled optional @lydell/node-pty dependency is 1.1.0. Set GEMINI_BUNDLE below to the installed @google/gemini-cli/bundle directory, save as repro.mjs, and run with node repro.mjs.

import path from 'node:path';
import {pathToFileURL} from 'node:url';
import {spawnSync} from 'node:child_process';
const bundle = process.env.GEMINI_BUNDLE;
// The actual 0.62.0 CLI path is gemini.js -> gemini-J6X2T2ZI.js -> this chunk.
// Update this filename from the CLI entry-point imports for a different build.
const entry = 'chunk-MLY4WQFO.js';
const {ShellExecutionService} = await import(pathToFileURL(path.join(bundle, entry)));
const count = () => spawnSync('lsof', ['-nP', '-a', '-p', String(process.pid), '/dev/ptmx'], {encoding:'utf8'}).stdout.split('\n').filter(x => x.includes('/dev/ptmx')).length;
for (const usePty of [true, false]) {
  const before = count();
  const methods = [];
  for (let i=0; i<6; i++) {
    const handle = await ShellExecutionService.execute(
      'printf pty-probe', '/tmp', () => {},
      new AbortController().signal, usePty,
      {env:{PATH:process.env.PATH, HOME:process.env.HOME},
       sanitizationConfig:{allowedEnvironmentVariables:[], blockedEnvironmentVariables:[]}}
    );
    const result = await handle.result;
    methods.push([result.executionMethod, result.exitCode, result.output]);
  }
  await new Promise(resolve => setTimeout(resolve, 5000));
  console.log({usePty, before, after:count(), methods});
}
process.exit(0);

Observed:

• usePty=true: before=0, after=6; all six methods lydell-node-pty, exit 0, output pty-probe.
• usePty=false: before=6, after=6; all six methods child_process, exit 0, output pty-probe.
• Counts remain unchanged after a five-second grace period following each command batch, ruling out short delayed-close scheduling. Normal probe exit releases its own retained handles.

Expected: completed commands should release their native PTY master while the agent process remains alive.

Mitigation verified

Launching with GEMINI_PTY_INFO=child_process makes the installed getPty() return null. Six commands routed through the real shell service then use child_process and retain zero masters, including when the service receives shouldUseNodePty=true. This avoids interactive shell PTYs but preserves the outer agent TUI. It does not recover already-retained handles in running agents.

The reproduction identifies the native shell path as the causal boundary; I have not yet isolated the precise native descriptor ownership defect. It does not depend on background commands, aborted commands, or MCP subprocess cleanup.