[dependencies]
nurlbox = "^0.2.0"

nurlbox — the Swiss Army knife of NURL

One binary, many utilities. nurlbox is a busybox-shaped multi-call executable: it decides which utility it is from the name it was invoked under, so a directory of symlinks pointing at it is a complete userland.

nurlbox                     # list the applets
nurlbox cat file            # run one by name
ln -s nurlbox /usr/bin/cat  # …or install it as itself
cat file

Everything is pure NURL over the shipped standard library. There is no shelling out, no coreutils underneath, and nothing to link beyond libc — which is also why the same source builds for Linux, macOS, Windows and wasm32-wasi, and boots as its own kernel on the unikernel target.

The applets

Group	Applets
shell	sh — quoting, $VAR / ${VAR:-…} / ${VAR#pat}, $(…), ` … , $((…)), globbing, pipelines, redirections, heredocs, && / ||, if / while / until / for / case, functions, and the builtins (cd export unset shift set read eval . source return break continue local type`)
text	cat echo head tail wc seq yes tac rev nl cut tr sort uniq tee fold expand unexpand paste comm shuf dos2unix unix2dos
search & edit	grep egrep fgrep sed find xargs
files	ls stat du df cp mv rm mkdir rmdir ln touch chmod readlink realpath truncate mktemp split
binary	od hexdump xxd cmp dd strings
digests	md5sum sha1sum sha256sum sha512sum cksum crc32 sum base64
archives	tar gzip gunzip zcat
processes	ps kill killall pidof free uptime mount
shell plumbing	test [ expr sleep usleep true false env printenv printf which factor
system	pwd basename dirname uname arch hostname whoami id groups logname nproc date sync clear tty

nurlbox --install [-s] DIR fills a directory with one entry per applet — hard links by default, symlinks with -s — which is how one file becomes a userland.

The shell

sh is the reason the rest of it is useful, and it makes one decision worth knowing about: an applet runs in-process. When a command names one of nurlbox's own, the shell calls it directly instead of exec'ing itself. That is busybox's standalone-shell trick, and here it is what makes the shell work on a machine with no fork at all.

On the unikernel that is not a detail — it is the whole thing:

$ NURL_APPEND='args="sh script.sh"' unikernel/run_qemu.sh nurlbox.elf

runs a script with globbing, loops, functions, pipelines, redirections and command substitution inside a guest that has exactly one address space and no processes. A pipeline there is not concurrent — each stage runs to completion into a temporary file and the next reads it — and the shell says so rather than pretending, exactly as it says so when a machine has nowhere writable to put that file.

What "clone" means here

The specification is the original, and the test suite says so: every case in tests/cases.sh runs the same command line through nurlbox and through the system busybox, and the two must agree on stdout, on the bytes, and on the exit status. Where an applet mutates the filesystem the comparison is the resulting tree — names, contents and permission bits — because that is a cp's real output, not what it printed.

./tests/nurlbox_test.sh          # ~260 differential cases
BUSYBOX=/path/to/busybox ./tests/nurlbox_test.sh

Three deliberate divergences, each because the original takes a shortcut a modern tool should not:

• Byte fidelity. A file whose last line has no newline does not grow

one, in any applet. busybox's cat -n and sed both append one.

• CRLF survives. head, cat, sed and the rest copy the line

terminator the input carried, rather than normalising it to \n.

• No trailing whitespace. ls -lh prints total 12K, not

total 12K····.

Where busybox does not implement an option at all (cat -E, cat -s, nl -n, cksum), the reference is GNU coreutils, and the suite compares against that instead. And find sorts each directory it reads, so a run is reproducible; the original emits in whatever order the filesystem answered.

What it needed from the language

The point of writing a userland is that it finds the holes. These were fixed where they actually were, not worked around here:

Gap	Fix
No stat(2) at all	FileStat + fs_stat / fs_lstat / fs_fstat, mode strings, fs_set_times, fs_user_name — stdlib/std/fs.nu over a new fixed-layout runtime thunk
No filesystem statistics	FsStat + fs_statfs — what df asks
No way to enumerate the environment	env_count / env_entry / env_list — stdlib/ext/env.nu
No uname, host name or processor count	stdlib/std/sysinfo.nu
Time was UTC-only	tz_offset / tz_name / time_local — stdlib/std/time.nu, plus a dozen more strftime directives
bufio could not read a line verbatim	bufreader_read_line_raw, which keeps the terminator
Regular expressions had no capture groups	a Pike VM with capture slots in stdlib/ext/regex.nu; regex_find_caps, regex_ngroups, regex_expand — sed's \1 is why
The wildcard matcher was private to fs_glob	fs_match / fs_match_glob
path_dirname "foo" answered "", path_basename "/a/b/" answered "", path_join "/" "x" answered "//x"	POSIX semantics in stdlib/std/path.nu
file_delete could not delete a dangling symlink	it stopped probing with access(2) first
mkdir ignored the umask	nurl_dir_create passes 0777 and lets the kernel subtract
fs_tempfile "/" produced //tmp.XXXXXX	one separator, not two
An owned temporary passed to a copying constructor leaked	mem_consumer_copy_safe in the compiler now covers string_from, the path_* family and nurl_eprint — ( string_from ( nurl_str_slice … ) ) no longer leaks its inner buffer, in any program

And on the freestanding side, so that all of the above works in a guest:

Gap	Fix
nolibc had no realpath, fdopen, link, symlink, readlink, chmod, getcwd, getuid/getgid/geteuid/getegid, getgroups, uname, sysconf, utimensat, localtime_r, statvfs	each one implemented where the target can answer, and refusing out loud where it cannot
The guest VFS had no stat and no directory listing	vfs_stat_path + a union directory handle over the baked-in archive and the disk, with . / .. folded
The guest had no working directory	one, with every path-taking entry point resolving against it — which is what cd is
The guest could not redirect a descriptor	a redirection table for fds 0–2 behind dup / dup2 / fcntl(F_DUPFD), so >, 2>&1, $(…) and pipelines work there

Memory

Leak-clean under AddressSanitizer with detect_leaks=1 across every applet — the manual-handle contract (String, Vec) is honoured on every path, including the error paths.