1.6.0: Elixir's CLI consumes application arguments and halts the VM

[joshrotenberg]

Two regressions in 1.6.0, same root cause. Verified with a 12-line application whose only dependency is burrito.

Context

1.5.x can no longer build natively on macOS. macos-latest is macos-26-arm64, and Zig 0.15.2 cannot link against the macOS 26 SDK:

error: undefined symbol: __availability_version_check
error: undefined symbol: _dispatch_queue_create
error: undefined symbol: _arc4random_buf

1.6.0 moves to Zig 0.16 and builds there. So neither released version both builds on current macOS and runs correctly.

1. Application arguments are consumed

$ probe hello world
No file named hello

$ probe --version
Erlang/OTP 29 [erts-17.0.3] ...
Elixir 1.20.2 (compiled with Erlang/OTP 29)

$ probe --help
--help : Standalone options can't be combined with other options

The application receives its arguments correctly; printing from start/2 shows :init.get_plain_arguments/0 returning ["hello", "world"]. They are consumed afterwards.

2. The VM halts, killing long-running applications

Independent of arguments. This application never halts:

def start(_type, _args) do
  spawn(fn -> Process.sleep(60_000) end)
  Supervisor.start_link([], strategy: :one_for_one, name: ArgvProbe.Sup)
end

Run with no arguments at all:

	result
1.6.0	exits after 1289ms, rc=0
1.5.0 semantics	stays up

Servers, Phoenix applications, and anything else expecting to keep running exit immediately after boot.

Cause

src/erlang_launcher.zig boots the release with -s elixir start_cli and appends the application's arguments after -extra.

In 1.5.0 that flag was a single argv element, "-s elixir start_cli". erl does not split argv elements, so it was never recognised and start_cli never ran. 1.6.0 splits it into "-s", "elixir", "start_cli" (lines 59-61), so it now runs.

Elixir's CLI then does what it is supposed to do: claims --version and --help, treats the first remaining plain argument as a script path, and halts the VM when it finishes.

$ elixir nosuchthing
No file named nosuchthing

$ elixir -- foo bar        # exits in 149ms
$ elixir --no-halt -- foo bar    # stays up

Fix

Emit --no-halt and -- after -extra, and strip them back off in Burrito.Util.Args. --no-halt restores the 1.5.0 lifetime, -- stops the CLI claiming arguments, and start_cli still runs, so whatever #225 fixed by splitting the flag is unaffected.

Reverting to the single joined element would also fix both, but only by making start_cli silently not run again.

Verified on macOS 26 arm64 and Linux arm64: arguments arrive intact including --version and --help, long-running applications stay up, exit codes propagate, maintenance subcommands work, and the EPIPE handling from #225 still terminates cleanly when piped into head.

Note System.argv/0 is populated only once start_cli runs, which is after Application.start/2, so applications still need Burrito.Util.Args.

PR follows.