//! A pardes launched inside a pardes hands its file to the outer one.
//!
//! Every top-level instance listens on `<dir>/pardes-<pid>.sock`, where `<dir>`
//! is `$XDG_RUNTIME_DIR` or, when the session has none, `~/.local/state/pardes`
//! created 0700. NOT /tmp: this socket takes a command line and runs it, and a
//! world-writable directory means both that somebody else can plant a listener
//! at a pid we are about to guess and that a file they planted under the sticky
//! bit cannot be unlinked, so bind fails and the feature goes quietly off.
//!
//! An instance that finds an ancestor process running the same executable
//! resolves its positional argument, writes ONE line — `Look /abs/path` — to
//! that ancestor's socket and exits silently; the outer pardes runs the line
//! through executeBuiltinLine and opens a pane for it. The wire format is a
//! builtin command line because that is a language pardes already speaks: no
//! serialization, nothing to version. The receive side still filters it down
//! to `Look `, because executeBuiltinLine dispatches ANY builtin and this
//! socket sits at a path anyone can derive from a pid — `Exec …` arriving here
//! is not something this protocol is allowed to say.
//!
//! Linux and darwin. The two differ in every primitive this needs and in none
//! of the design: /proc against libproc for the ancestor walk, SOCK_CLOEXEC
//! and accept4 against a plain socket plus an fcntl, and a `sun_path` of 108
//! bytes against one of 104 — which is why no buffer below spells a number,
//! they are all sized from the field itself. Anywhere else the walk returns
//! null and a pardes inside a pardes opens a second session, as before.
//!
//! macOS also has a third executable in the family: the app bundle. Its binary
//! is named `pardes`, like the tty frontend, wherever the bundle is installed.
//! Executable identity therefore comes from the family name rather than its
//! path — see samePardesExecutable.
const std = @import("std");
const builtin = @import("builtin");
const libc = std.c;

// std.c has getenv but neither setter; the tests below need both
extern "c" fn setenv(name: [*:0]const u8, value: [*:0]const u8, overwrite: c_int) c_int;
extern "c" fn unsetenv(name: [*:0]const u8) c_int;

/// THE SOCKET CONVENTIONS BELOW ARE SHARED, and the ones marked `pub` are
/// shared with src/detached/server.zig — a second unix socket in the same
/// per-user directory, under a different name (`pardes-detached-<name>.sock`
/// rather than `pardes-<pid>.sock`). They were copied into that file when it
/// landed; one directory vetted by two different predicates is exactly the
/// divergence the reasoning here is meant to prevent, so there is one of each.
pub const darwin = switch (builtin.os.tag) {
    .macos, .ios, .tvos, .watchos, .visionos => true,
    else => false,
};

/// This module is only as portable as its two ingredients: a way to name the
/// executable and parent of an arbitrary pid, and unix sockets. The detached
/// transport needs the second alone, and the same answer.
pub const supported = builtin.os.tag == .linux or darwin;

/// `sun_path` is 108 bytes on linux and 104 on darwin, and it is the hard
/// limit on this whole feature: a path that does not fit is not a socket
/// address, it is a truncated one pointing somewhere else. Taken from the
/// struct so that the buffers, the fit checks and the memcpy below cannot
/// disagree with the kernel or with each other.
pub const sun_path_len = @typeInfo(@FieldType(libc.sockaddr.un, "path")).array.len;

/// libproc, darwin's answer to /proc. `proc_pidpath` is readlink of
/// `/proc/<pid>/exe`; `PROC_PIDTBSDINFO` carries the parent pid that linux
/// spells `PPid:`. Both are same-uid readable, which is the only permission
/// an ancestor walk through one's own processes needs.
const PROC_PIDTBSDINFO: c_int = 3;
const proc_bsdinfo = extern struct {
    flags: u32,
    status: u32,
    xstatus: u32,
    pid: u32,
    ppid: u32,
    /// uids, gids, comm, name, the tty and the start time: filled by the
    /// kernel and unread here, but the call fails unless the buffer is the
    /// whole 136-byte record.
    rest: [116]u8,
};
extern "c" fn proc_pidpath(pid: c_int, buffer: *anyopaque, buffersize: u32) c_int;
extern "c" fn proc_pidinfo(pid: c_int, flavor: c_int, arg: u64, buffer: *anyopaque, buffersize: c_int) c_int;

/// Linux opens sockets CLOEXEC in one call; darwin has to set it afterwards.
/// The gap is a race only against a fork on another thread, and every caller
/// is past that: `listen` runs before the first pane exists, `acceptLine` runs
/// on a thread of its own long after spawning has settled, and the detached
/// session forks nothing at all (its ptys live in its frontends).
///
/// CLOEXEC still matters for both: a `--detach` session is long-lived, and an
/// inherited listener would keep its socket bound long after it ended — the
/// same shape as the inherited lock fd that once held a flock forever.
pub fn setCloexec(fd: c_int) void {
    const FD_CLOEXEC: c_int = 1;
    _ = libc.fcntl(fd, libc.F.SETFD, FD_CLOEXEC);
}

/// The longest command line this protocol carries or accepts. `Look ` plus a
/// PATH_MAX path fits with room over; anything longer cannot have come from
/// the client and is dropped rather than truncated into a different command.
pub const max_line = 4200;

/// Where the sockets live. `$XDG_RUNTIME_DIR` first — a per-user 0700 tmpfs
/// the login session already cleans up — else `~/.local/state/pardes`, which
/// is per-user for the same reason a home directory is. NEVER /tmp: these
/// sockets take a command line, or keystrokes into a live editor. Asked by the
/// client (to derive the path), by the listener (to create and vet it), by the
/// sweeper (to scan it) and by the detached transport (all three, for its own
/// name), so it is written once.
pub fn socketDir(buf: *[sun_path_len:0]u8) ?[:0]const u8 {
    if (libc.getenv("XDG_RUNTIME_DIR")) |x|
        return std.fmt.bufPrintSentinel(buf, "{s}", .{std.mem.span(x)}, 0) catch null;
    const home = libc.getenv("HOME") orelse return null;
    return std.fmt.bufPrintSentinel(buf, "{s}/.local/state/pardes", .{std.mem.span(home)}, 0) catch null;
}

/// `<dir>/pardes-<pid>.sock`. `<pid>` is the LISTENING instance's own pid, so
/// two pardes never collide and a nested child derives the exact path from the
/// ancestor pid its tree walk found. The buffer is sun_path-sized: a longer
/// path is not a socket address at all.
pub fn socketPath(buf: *[sun_path_len]u8, pid: libc.pid_t) ?[:0]const u8 {
    var dir_buf: [sun_path_len:0]u8 = undefined;
    const dir = socketDir(&dir_buf) orelse return null;
    // unsigned: {d} prints a leading '+' for a positive SIGNED int
    return std.fmt.bufPrintSentinel(buf, "{s}/pardes-{d}.sock", .{ dir, @as(u32, @intCast(pid)) }, 0) catch null;
}

/// Normalize the kernel suffix left on a running executable after its file is
/// replaced. `zig build` does this routinely while an outer session is live.
fn stripDeleted(link: []const u8) []const u8 {
    const suffix = " (deleted)";
    return if (std.mem.endsWith(u8, link, suffix)) link[0 .. link.len - suffix.len] else link;
}

/// The tty, SDL and macOS builds are sibling frontends of the same program.
/// Their installed names differ only by `-gui` (and, for cross builds, share
/// the same `-os-arch` tail), or not at all when one of them is the app bundle
/// — so any of them must recognise any other as an outer pardes. Paths are
/// deliberately ignored: the GUI may be installed system-wide while the tty
/// frontend is installed in the user's bin directory.
fn samePardesExecutable(a_raw: []const u8, b_raw: []const u8) bool {
    const a = stripDeleted(a_raw);
    const b = stripDeleted(b_raw);
    return sameFamily(std.fs.path.basename(a), std.fs.path.basename(b));
}

/// What is left of a family name after the frontend part: `` for `pardes` and
/// `pardes-gui`, `-linux-aarch64` for the cross-built spellings of both. Null
/// when the name is not in the family at all — `not-pardes`, `pardesfoo`, and
/// helper binaries such as `pardes-snap` are other programs.
fn familyTail(name: []const u8) ?[]const u8 {
    const rest = if (std.mem.startsWith(u8, name, "pardes-gui"))
        name["pardes-gui".len..]
    else if (std.mem.startsWith(u8, name, "pardes"))
        name["pardes".len..]
    else
        return null;
    if (rest.len == 0) return rest;
    // Build names have exactly `-os-arch` after the frontend. Validating both
    // fields keeps sibling installs flexible without mistaking pardes-snap,
    // pardes-perf, and the other helper executables for editor frontends.
    if (rest[0] != '-') return null;
    var fields = std.mem.splitScalar(u8, rest[1..], '-');
    const os = fields.next() orelse return null;
    const arch = fields.next() orelse return null;
    if (fields.next() != null) return null;
    if (std.meta.stringToEnum(std.Target.Os.Tag, os) == null) return null;
    if (std.meta.stringToEnum(std.Target.Cpu.Arch, arch) == null) return null;
    return rest;
}

/// Two executable names in the same family. The tails have to agree — a linux
/// binary and an x86_64 one are two builds — unless one of them has no tail at
/// all, which is the untagged name the default build and, unavoidably, the app
/// bundle both produce: CFBundleExecutable is a fixed string, so the bundled
/// copy of `pardes-macos-aarch64` is called `pardes` and nothing in the name
/// records what it was. A foreign-arch ancestor cannot be running here.
fn sameFamily(a: []const u8, b: []const u8) bool {
    const a_tail = familyTail(a) orelse return false;
    const b_tail = familyTail(b) orelse return false;
    if (std.mem.eql(u8, a, b)) return true;
    return a_tail.len == 0 or b_tail.len == 0 or std.mem.eql(u8, a_tail, b_tail);
}

/// The `PPid:` field of a /proc/<pid>/status blob. Deliberately NOT field 4 of
/// /proc/<pid>/stat: that field is positional after `comm`, and a comm may
/// contain spaces and parentheses — a process named `sh (a b)` shifts every
/// field after it and the parse silently reads the wrong number.
fn parsePPid(status: []const u8) ?libc.pid_t {
    var lines = std.mem.splitScalar(u8, status, '\n');
    while (lines.next()) |line| {
        if (!std.mem.startsWith(u8, line, "PPid:")) continue;
        return std.fmt.parseInt(libc.pid_t, std.mem.trim(u8, line["PPid:".len..], " \t\r"), 10) catch null;
    }
    return null;
}

/// The pid in a `pardes-<pid>.sock` filename, for the startup sweep. Strictly
/// digits: parseInt alone would take `pardes-+7.sock` and `pardes--7.sock`,
/// and the sweep unlinks what this answers about.
fn sweepPid(name: []const u8) ?libc.pid_t {
    if (!std.mem.startsWith(u8, name, "pardes-") or !std.mem.endsWith(u8, name, ".sock")) return null;
    const digits = name["pardes-".len .. name.len - ".sock".len];
    if (digits.len == 0) return null;
    for (digits) |ch| if (!std.ascii.isDigit(ch)) return null;
    return std.fmt.parseInt(libc.pid_t, digits, 10) catch null;
}

/// Name the executable behind a pid, the way this OS spells it.
fn exeOf(pid: libc.pid_t, buf: *[4096]u8) ?[]const u8 {
    switch (builtin.os.tag) {
        .linux => {
            var name: [64:0]u8 = undefined;
            const link = std.fmt.bufPrintSentinel(&name, "/proc/{d}/exe", .{@as(u32, @intCast(pid))}, 0) catch return null;
            const n = libc.readlink(link, buf, buf.len);
            if (n <= 0) return null;
            return buf[0..@intCast(n)];
        },
        else => {
            if (comptime !darwin) return null;
            // Documented to want a PROC_PIDPATHINFO_MAXSIZE buffer, which is
            // exactly this one, and to return the length it wrote.
            const n = proc_pidpath(pid, buf, @intCast(buf.len));
            if (n <= 0) return null;
            return buf[0..@intCast(n)];
        },
    }
}

/// ...and its parent.
fn parentOf(pid: libc.pid_t) ?libc.pid_t {
    switch (builtin.os.tag) {
        .linux => {
            var name: [64:0]u8 = undefined;
            var buf: [4096]u8 = undefined;
            const status = std.fmt.bufPrintSentinel(&name, "/proc/{d}/status", .{@as(u32, @intCast(pid))}, 0) catch return null;
            const fd = libc.open(status, .{ .ACCMODE = .RDONLY });
            if (fd < 0) return null;
            const got = libc.read(fd, &buf, buf.len);
            _ = libc.close(fd);
            if (got <= 0) return null;
            return parsePPid(buf[0..@intCast(got)]);
        },
        else => {
            if (comptime !darwin) return null;
            var info: proc_bsdinfo = undefined;
            const n = proc_pidinfo(pid, PROC_PIDTBSDINFO, 0, &info, @sizeOf(proc_bsdinfo));
            // A short answer means the record this was compiled against is not
            // the one the kernel filled, and `ppid` is then some other field.
            if (n < @as(c_int, @sizeOf(proc_bsdinfo))) return null;
            return @intCast(info.ppid);
        },
    }
}

/// The pid of the nearest ancestor running a pardes executable, or null.
/// Identity is that ancestor's executable path against our own; the tty, SDL
/// and app-bundle siblings also match when they were installed together. A
/// name alone would call every unrelated `pardes` ancestor an outer instance.
/// The hop cap is not for the process tree, which cannot loop, but because the
/// walk is driven by numbers read out of the kernel and should not be able to
/// spin on a surprising one.
pub fn outer() ?libc.pid_t {
    if (comptime !supported) return null;
    var self_buf: [4096]u8 = undefined;
    const self_exe = exeOf(libc.getpid(), &self_buf) orelse return null;
    // A process harness may deliberately launch a fresh top-level pardes from
    // inside another one. Its pid is a process-tree boundary, not an opt-out
    // for the new session itself: pane shells below the child still detect it.
    // This is what lets the snapshot harness exercise nested launches while
    // the harness happens to be running in a real pardes pane.
    const boundary = if (libc.getenv("PARDES_NESTED_BOUNDARY_PID")) |raw|
        std.fmt.parseInt(libc.pid_t, std.mem.span(raw), 10) catch 0
    else
        0;
    var pid = libc.getppid();
    var hops: usize = 0;
    while (pid > 1 and hops < 64) : (hops += 1) {
        if (pid == boundary) return null;
        var buf: [4096]u8 = undefined;
        if (exeOf(pid, &buf)) |exe| if (samePardesExecutable(exe, self_exe)) return pid;
        pid = parentOf(pid) orelse return null;
    }
    return null;
}

/// Hand `Look <path>[:<line>]` to the pardes listening as `pid` and say
/// whether it landed. False for every failure — no socket file, nobody
/// accepting, a path that does not fit — because an outer instance that
/// cannot be reached (an older build, a stale path) must never cost the
/// caller its own launch. Writes and returns: the answer is a pane appearing
/// on someone else's screen, and there is nothing to wait for.
pub fn sendLook(pid: libc.pid_t, path: []const u8, line: usize) bool {
    if (comptime !supported) return false;
    // The protocol is one line, so a path with a line break IN it says
    // something else entirely: `we\nird.txt` arrived as `Look .../we` and the
    // outer instance opened a different file that happened to exist. \r goes
    // too — the receive side trims a trailing one. Unsendable, not escaped:
    // the caller falls through and opens the file in its own session.
    if (std.mem.indexOfAny(u8, path, "\r\n") != null) return false;
    var cmd_buf: [max_line]u8 = undefined;
    const cmd = (if (line > 0)
        std.fmt.bufPrint(&cmd_buf, "Look {s}:{d}\n", .{ path, line })
    else
        std.fmt.bufPrint(&cmd_buf, "Look {s}\n", .{path})) catch return false;

    // sun_path-sized by construction, so `sock` cannot be longer than the
    // field it is about to be copied into — socketPath returns null instead.
    var path_buf: [sun_path_len]u8 = undefined;
    const sock = socketPath(&path_buf, pid) orelse return false;
    var addr: libc.sockaddr.un = .{ .path = @splat(0) };
    @memcpy(addr.path[0 .. sock.len + 1], sock[0 .. sock.len + 1]);
    const fd = libc.socket(libc.AF.UNIX, libc.SOCK.STREAM, 0);
    if (fd < 0) return false;
    setCloexec(fd);
    defer _ = libc.close(fd);
    if (libc.connect(fd, @ptrCast(&addr), @sizeOf(@TypeOf(addr))) != 0) return false;
    var off: usize = 0;
    while (off < cmd.len) {
        const n = libc.write(fd, cmd.ptr + off, cmd.len - off);
        if (n < 0) {
            if (libc.errno(n) == .INTR) continue;
            return false;
        }
        if (n == 0) return false;
        off += @intCast(n);
    }
    return true;
}

/// The two things `ensureSocketDir` has to know about a path, from whichever
/// call the platform actually offers. Darwin has fstatat and no statx; on
/// linux std.c.fstatat is `void` — glibc hides it behind a versioned symbol
/// std cannot name — so linux asks statx for the same fields. Both spellings
/// refuse to follow a symlink, which is the point of asking.
///
/// `pub` for the detached transport, which vets the same directory and also
/// vets the SOCKET FILE with it (src/detached/server.zig `vetted`): `mode`
/// carries the type bits, so one call answers "is this a socket, ours, and
/// private" as well as it answers it for a directory.
pub const DirFacts = struct { mode: u32, uid: libc.uid_t };

pub fn statNoFollow(path: [:0]const u8) ?DirFacts {
    if (comptime darwin) {
        var st: libc.Stat = undefined;
        if (libc.fstatat(libc.AT.FDCWD, path, &st, libc.AT.SYMLINK_NOFOLLOW) != 0) return null;
        return .{ .mode = st.mode, .uid = st.uid };
    } else {
        const linux = std.os.linux;
        var stx: linux.Statx = undefined;
        const want: linux.STATX = .{ .TYPE = true, .MODE = true, .UID = true };
        if (libc.statx(linux.AT.FDCWD, path, linux.AT.SYMLINK_NOFOLLOW, want, &stx) != 0) return null;
        return .{ .mode = stx.mode, .uid = stx.uid };
    }
}

/// Create the socket directory if it is missing and refuse it unless it is a
/// directory WE own with nothing granted to group or other. A planted path is
/// the whole attack on a socket that runs commands — or, for the detached
/// transport that shares this, on one that carries keystrokes into a live
/// editor — and $XDG_RUNTIME_DIR passes this untouched (the login session
/// already makes it 0700).
pub fn ensureSocketDir(dir: [:0]const u8) bool {
    // mkdir -p, because the HOME branch is three levels deep and a machine
    // without ~/.local/state would otherwise switch the feature off in
    // silence. Under $XDG_RUNTIME_DIR every prefix already exists and simply
    // EEXISTs, which is the ordinary case for the leaf too.
    var partial: [sun_path_len:0]u8 = undefined;
    @memcpy(partial[0 .. dir.len + 1], dir[0 .. dir.len + 1]);
    for (1..dir.len) |i| {
        if (dir[i] != '/') continue;
        partial[i] = 0;
        _ = libc.mkdir(partial[0..i :0], 0o700);
        partial[i] = '/';
    }
    _ = libc.mkdir(dir, 0o700);
    // A symlink where the directory should be is exactly the plant this
    // guards against, so the stat above it does not follow one.
    const st = statNoFollow(dir) orelse return false;
    const IFMT: u32 = 0o170000;
    const IFDIR: u32 = 0o040000;
    if (st.mode & IFMT != IFDIR) return false;
    if (st.uid != libc.getuid()) return false;
    return st.mode & 0o077 == 0;
}

/// Unlink the socket files of pardes processes that are gone. A pardes killed
/// rather than quit runs no defer, so its file outlives it; harmless by
/// construction (bind unlinks first, a client's connect is refused) but it is
/// our own litter and the snapshot suite alone leaves ~90 behind per run.
/// Bounded: one readdir of a directory only we write to, one kill(0) each.
fn sweep(dir: [:0]const u8) void {
    const d = libc.opendir(dir) orelse return;
    defer _ = libc.closedir(d);
    const me = libc.getpid();
    while (libc.readdir(d)) |ent| {
        const pid = sweepPid(std.mem.sliceTo(&ent.name, 0)) orelse continue;
        if (pid == me) continue;
        // 0 = alive; EPERM = alive and someone else's. Only ESRCH is a corpse.
        const rc = libc.kill(pid, @enumFromInt(0));
        if (rc == 0 or libc.errno(rc) != .SRCH) continue;
        var pbuf: [sun_path_len]u8 = undefined;
        _ = libc.unlink(socketPath(&pbuf, pid) orelse continue);
    }
}

/// Bind and listen so nested instances can find us; -1 if anything fails, and
/// a pardes without a socket is simply one whose children open their own UI.
/// The path is always this process's own, so nobody outside holds a buffer of
/// it — the shells each kept one and passed it back to be unlinked, which is a
/// way for the two spellings to go out of step and for no other reason.
///
/// CLOEXEC matters more here than on any other fd in the program: pane shells
/// are forked with forkpty and inherit everything open, and an orphaned bash
/// holding this one would keep the socket bound long after we exit — the same
/// shape as the inherited lock fd that once held a flock forever.
pub fn listen() c_int {
    if (comptime !supported) return -1;
    var dir_buf: [sun_path_len:0]u8 = undefined;
    const dir = socketDir(&dir_buf) orelse return -1;
    if (!ensureSocketDir(dir)) return -1;
    sweep(dir);
    // Fits by construction: socketPath writes into a sun_path-sized buffer and
    // returns null rather than a truncated address.
    var path_buf: [sun_path_len]u8 = undefined;
    const path = socketPath(&path_buf, libc.getpid()) orelse return -1;
    var addr: libc.sockaddr.un = .{ .path = @splat(0) };
    @memcpy(addr.path[0 .. path.len + 1], path[0 .. path.len + 1]);
    const fd = libc.socket(libc.AF.UNIX, libc.SOCK.STREAM, 0);
    if (fd < 0) return -1;
    setCloexec(fd);
    _ = libc.unlink(path); // pid reuse: a dead pardes' file would EADDRINUSE forever
    if (libc.bind(fd, @ptrCast(&addr), @sizeOf(@TypeOf(addr))) != 0) {
        _ = libc.close(fd);
        return -1;
    }
    // Owner-only, and BEFORE listen(2), which is the moment anyone could
    // connect: the directory is already private, this is the second wall.
    _ = libc.chmod(path, 0o600);
    if (libc.listen(fd, 8) != 0) {
        _ = libc.close(fd);
        return -1;
    }
    return fd;
}

/// Close the listener and take its file away. Guarded on the fd rather than on
/// the path, so a bind that FAILED cannot unlink a path this process never
/// created; anything else is a no-op, which is what --nested and every
/// unsupported build hand it.
pub fn unlisten(fd: c_int) void {
    if (fd < 0) return;
    _ = libc.close(fd);
    var path_buf: [sun_path_len]u8 = undefined;
    if (socketPath(&path_buf, libc.getpid())) |path| _ = libc.unlink(path);
}

/// Block until a nested instance sends a `Look` line, and return it inside
/// `buf`. Null only when the listening fd itself is gone — teardown closed it,
/// or it was never a socket — because anything else (EMFILE, ECONNABORTED)
/// would otherwise kill the listener thread for the life of the process while
/// the socket stayed bound, and every later launch would exit 0 having done
/// nothing. Every accepted connection is CLOEXEC for the reason the listener
/// is.
pub fn acceptLine(fd: c_int, buf: []u8) ?[]const u8 {
    if (comptime !supported) return null;
    while (true) {
        const conn = libc.accept(fd, null, null);
        if (conn < 0) {
            switch (libc.errno(conn)) {
                .INTR => continue,
                // the fd went away or never was one: nothing will ever arrive
                .BADF, .INVAL, .NOTSOCK => return null,
                // transient. Sleep first: EMFILE persists until some other fd
                // is freed, and a bare `continue` would spin a core on it.
                else => {
                    var ts: libc.timespec = .{ .sec = 0, .nsec = 100 * std.time.ns_per_ms };
                    _ = libc.nanosleep(&ts, null);
                    continue;
                },
            }
        }
        defer _ = libc.close(conn);
        setCloexec(conn);
        // A peer that connects and says nothing must not hold the listener:
        // this is a serial accept loop, and one silent connection used to
        // block every later launch until it let go. The client writes its one
        // short line immediately, so a second is already generous.
        const tv: libc.timeval = .{ .sec = 1, .usec = 0 };
        _ = libc.setsockopt(conn, libc.SOL.SOCKET, libc.SO.RCVTIMEO, &tv, @sizeOf(libc.timeval));
        var len: usize = 0;
        while (len < buf.len) {
            const n = libc.read(conn, buf.ptr + len, buf.len - len);
            if (n < 0 and libc.errno(n) == .INTR) continue;
            if (n <= 0) break; // EOF, or the receive timeout expired
            len += @intCast(n);
            if (std.mem.indexOfScalar(u8, buf[0..len], '\n') != null) break;
        }
        const end = std.mem.indexOfScalar(u8, buf[0..len], '\n') orelse len;
        // a full buffer with no newline is an overlong line: drop it whole
        // rather than run its truncation as some other command
        if (end == buf.len) continue;
        const line = std.mem.trimEnd(u8, buf[0..end], "\r");
        // one verb (see the file header): this socket may open things, and
        // that is all it may do
        if (!std.mem.startsWith(u8, line, "Look ")) continue;
        return line;
    }
}

test "socket path: XDG first, then a private dir under HOME, never /tmp" {
    var buf: [sun_path_len]u8 = undefined;
    // The environment is process-wide and every test in this binary shares it.
    // The last case below reaches the "no directory at all" branch by blanking
    // both variables, and without this every later test ran without a HOME.
    var xdg_buf: [4096:0]u8 = undefined;
    var home_buf: [4096:0]u8 = undefined;
    const xdg0 = if (libc.getenv("XDG_RUNTIME_DIR")) |v| std.fmt.bufPrintSentinel(&xdg_buf, "{s}", .{std.mem.span(v)}, 0) catch null else null;
    const home0 = if (libc.getenv("HOME")) |v| std.fmt.bufPrintSentinel(&home_buf, "{s}", .{std.mem.span(v)}, 0) catch null else null;
    defer {
        if (xdg0) |v| {
            _ = setenv("XDG_RUNTIME_DIR", v, 1);
        } else _ = unsetenv("XDG_RUNTIME_DIR");
        if (home0) |v| {
            _ = setenv("HOME", v, 1);
        } else _ = unsetenv("HOME");
    }
    _ = setenv("XDG_RUNTIME_DIR", "/run/user/1000", 1);
    try std.testing.expectEqualStrings("/run/user/1000/pardes-4242.sock", socketPath(&buf, 4242).?);
    _ = unsetenv("XDG_RUNTIME_DIR");
    _ = setenv("HOME", "/home/who", 1);
    try std.testing.expectEqualStrings("/home/who/.local/state/pardes/pardes-4242.sock", socketPath(&buf, 4242).?);
    // sun_path holds the NUL, so a directory that fills it has no socket
    // address at all — say so instead of binding a truncated one. Sized from
    // the field: the limit is 108 on linux and 104 on darwin, and a literal
    // here would test nothing on whichever platform it was not written for.
    _ = setenv("XDG_RUNTIME_DIR", "/" ++ ("x" ** (sun_path_len - 8)), 1);
    try std.testing.expect(socketPath(&buf, 4242) == null);
    _ = unsetenv("XDG_RUNTIME_DIR");
    _ = unsetenv("HOME");
    try std.testing.expect(socketPath(&buf, 4242) == null);
}

test "a rebuilt binary still matches its own running instance" {
    // `zig build` under a live pardes: the outer's exe link gains the suffix,
    // the new process's does not, and before this the two stopped comparing
    // equal — every nested launch opened a second UI.
    try std.testing.expectEqualStrings("/usr/bin/pardes", stripDeleted("/usr/bin/pardes (deleted)"));
    try std.testing.expectEqualStrings("/usr/bin/pardes", stripDeleted("/usr/bin/pardes"));
    try std.testing.expectEqualStrings("", stripDeleted(" (deleted)"));
    // only a SUFFIX, and only the whole one
    try std.testing.expectEqualStrings("/x (deleted) y", stripDeleted("/x (deleted) y"));
    try std.testing.expectEqualStrings("/x (delete)", stripDeleted("/x (delete)"));
}

test "tty and GUI sibling executables recognise each other" {
    try std.testing.expect(samePardesExecutable(
        "/work/zig-out/bin/pardes",
        "/work/zig-out/bin/pardes-gui",
    ));
    try std.testing.expect(samePardesExecutable(
        "/work/zig-out/bin/pardes-linux-aarch64",
        "/work/zig-out/bin/pardes-gui-linux-aarch64 (deleted)",
    ));
    // Installation paths do not define the family. This is the ordinary
    // system-GUI/user-TTY pairing and the reason this comparison uses names.
    try std.testing.expect(samePardesExecutable(
        "/home/who/.local/bin/pardes",
        "/usr/bin/pardes-gui",
    ));
    try std.testing.expect(!samePardesExecutable(
        "/work/zig-out/bin/pardes-linux-aarch64",
        "/work/zig-out/bin/pardes-gui-linux-x86_64",
    ));
    try std.testing.expect(!samePardesExecutable(
        "/work/zig-out/bin/not-pardes",
        "/work/zig-out/bin/not-pardes-gui",
    ));
    try std.testing.expect(!samePardesExecutable(
        "/one/bin/not-pardes",
        "/two/bin/not-pardes",
    ));
    try std.testing.expect(!samePardesExecutable(
        "/work/zig-out/bin/pardes-snap",
        "/usr/bin/pardes",
    ));
}

test "the app bundle is in the same executable family" {
    // What `pardes foo.zig` typed into the bundle's own shell has to resolve:
    // the ancestor is zig-out/pardes.app/..., this process is zig-out/bin/...,
    // and nothing below zig-out is shared.
    try std.testing.expect(samePardesExecutable(
        "/work/zig-out/pardes.app/Contents/MacOS/pardes",
        "/work/zig-out/bin/pardes",
    ));
    // ...and the SDL sibling, which reaches it by the name rule instead.
    try std.testing.expect(samePardesExecutable(
        "/work/zig-out/pardes.app/Contents/MacOS/pardes",
        "/work/zig-out/bin/pardes-gui",
    ));
    // The case this machine actually produces: `zig build` installs the tty
    // binary under its os-arch tail, and the bundle carries the same build
    // under the one name CFBundleExecutable can spell.
    try std.testing.expect(samePardesExecutable(
        "/work/zig-out/pardes.app/Contents/MacOS/pardes",
        "/work/zig-out/bin/pardes-macos-aarch64",
    ));
    // Installation location does not matter here either.
    try std.testing.expect(samePardesExecutable(
        "/work/zig-out/pardes.app/Contents/MacOS/pardes",
        "/opt/zig-out/bin/pardes",
    ));
    try std.testing.expect(samePardesExecutable(
        "/work/zig-out/pardes/Contents/MacOS/pardes",
        "/work/zig-out/bin/pardes",
    ));
    // Nothing here may loosen the rule for two unrelated programs that merely
    // sit in a bin and a bundle of the same tree.
    try std.testing.expect(!samePardesExecutable(
        "/work/zig-out/other.app/Contents/MacOS/other",
        "/work/zig-out/bin/pardes",
    ));
}

test "the ancestor walk reads this process's own parent" {
    // The one thing a hand-written `struct proc_bsdinfo` gets wrong silently:
    // a field ordering that puts something else where ppid should be still
    // returns a plausible number. getppid knows the answer, so compare.
    //
    // Also the only check that libproc answers us at all — every caller of
    // outer() treats a failure as "no outer instance", which is exactly what a
    // permission problem would look like.
    if (comptime !supported) return error.SkipZigTest;
    try std.testing.expectEqual(libc.getppid(), parentOf(libc.getpid()).?);
    // ...and that the walk terminates rather than spinning on pid 1's parent.
    try std.testing.expect(parentOf(1) == null or parentOf(1).? <= 1);

    var buf: [4096]u8 = undefined;
    const exe = exeOf(libc.getpid(), &buf).?;
    try std.testing.expect(exe.len > 0);
    try std.testing.expect(exe[0] == '/');
    // The test binary is not a pardes, so the walk must come back empty rather
    // than matching some ancestor by accident.
    try std.testing.expect(outer() == null);
}

extern "c" fn mkdtemp(template: [*:0]u8) ?[*:0]u8;
extern "c" fn rmdir(path: [*:0]const u8) c_int;

test "a Look line survives the socket round trip" {
    // Everything the protocol actually does, against a real kernel: bind,
    // chmod, connect, write, accept, read, and the one-verb filter. The pure
    // functions above cannot see any of it, and every primitive here is
    // spelled differently on the two platforms this now supports.
    if (comptime !supported) return error.SkipZigTest;

    // A private directory of our own. Not the developer's real state dir: this
    // binds a socket named after a pid that is the TEST's, and sweep() unlinks
    // what it finds beside it.
    var tmpl: [64:0]u8 = undefined;
    _ = std.fmt.bufPrintSentinel(&tmpl, "/tmp/pardes-nested-XXXXXX", .{}, 0) catch unreachable;
    if (mkdtemp(&tmpl) == null) return error.SkipZigTest;
    const dir = std.mem.sliceTo(&tmpl, 0);
    defer _ = rmdir(tmpl[0..dir.len :0]);

    var xdg_buf: [4096:0]u8 = undefined;
    const xdg0 = if (libc.getenv("XDG_RUNTIME_DIR")) |v| std.fmt.bufPrintSentinel(&xdg_buf, "{s}", .{std.mem.span(v)}, 0) catch null else null;
    defer {
        if (xdg0) |v| {
            _ = setenv("XDG_RUNTIME_DIR", v, 1);
        } else _ = unsetenv("XDG_RUNTIME_DIR");
    }
    _ = setenv("XDG_RUNTIME_DIR", tmpl[0..dir.len :0], 1);

    const fd = listen();
    try std.testing.expect(fd >= 0);
    defer unlisten(fd);

    // Sent to our own pid, which is the pid listen() named the socket after.
    // The client closes as it returns, and the line is already queued, so the
    // single-threaded accept below finds a complete connection waiting — no
    // thread and no timeout needed to prove the protocol.
    try std.testing.expect(sendLook(libc.getpid(), "/etc/hosts", 42));
    var buf: [max_line]u8 = undefined;
    try std.testing.expectEqualStrings("Look /etc/hosts:42", acceptLine(fd, &buf).?);

    // ...and without a line number, which is the directory and image case.
    try std.testing.expect(sendLook(libc.getpid(), "/etc", 0));
    try std.testing.expectEqualStrings("Look /etc", acceptLine(fd, &buf).?);

    // The socket takes one verb. Anything else is dropped rather than run, so
    // the next Look is what comes back — proving the filter skipped it without
    // dropping the connection after it.
    try std.testing.expect(writeLine(libc.getpid(), "Exec rm -rf /\n"));
    try std.testing.expect(sendLook(libc.getpid(), "/etc/passwd", 0));
    try std.testing.expectEqualStrings("Look /etc/passwd", acceptLine(fd, &buf).?);

    // A path that cannot be one line is not escaped, it is refused.
    try std.testing.expect(!sendLook(libc.getpid(), "/etc/ho\nsts", 0));
}

/// sendLook with the framing bypassed, so a test can put something on the wire
/// that the client would never send.
fn writeLine(pid: libc.pid_t, line: []const u8) bool {
    var path_buf: [sun_path_len]u8 = undefined;
    const sock = socketPath(&path_buf, pid) orelse return false;
    var addr: libc.sockaddr.un = .{ .path = @splat(0) };
    @memcpy(addr.path[0 .. sock.len + 1], sock[0 .. sock.len + 1]);
    const fd = libc.socket(libc.AF.UNIX, libc.SOCK.STREAM, 0);
    if (fd < 0) return false;
    defer _ = libc.close(fd);
    if (libc.connect(fd, @ptrCast(&addr), @sizeOf(@TypeOf(addr))) != 0) return false;
    return libc.write(fd, line.ptr, line.len) == @as(isize, @intCast(line.len));
}

test "the sweep only recognises its own socket names" {
    try std.testing.expectEqual(@as(libc.pid_t, 7), sweepPid("pardes-7.sock").?);
    try std.testing.expectEqual(@as(libc.pid_t, 4194304), sweepPid("pardes-4194304.sock").?);
    try std.testing.expect(sweepPid("pardes-.sock") == null);
    try std.testing.expect(sweepPid("pardes-7.sockx") == null);
    try std.testing.expect(sweepPid("pardes-7") == null);
    try std.testing.expect(sweepPid("bus") == null);
    try std.testing.expect(sweepPid("pardes-osc133.bash") == null);
    // parseInt alone would take these, and the sweep UNLINKS what it answers
    try std.testing.expect(sweepPid("pardes-+7.sock") == null);
    try std.testing.expect(sweepPid("pardes--7.sock") == null);
    try std.testing.expect(sweepPid("pardes- 7.sock") == null);
}

test "PPid comes off the status field, not a comm-shifted stat line" {
    // the comm here contains a space AND parentheses — the exact shape that
    // breaks `field 4 of /proc/<pid>/stat`
    const status = "Name:\tsh (a b)\nUmask:\t0022\nState:\tS (sleeping)\n" ++
        "Tgid:\t1234\nNgid:\t0\nPid:\t1234\nPPid:\t991\nTracerPid:\t0\n";
    try std.testing.expectEqual(@as(libc.pid_t, 991), parsePPid(status).?);
    try std.testing.expectEqual(@as(libc.pid_t, 0), parsePPid("PPid:\t0\n").?);
    try std.testing.expect(parsePPid("Name:\tinit\nTracerPid:\t0\n") == null);
    try std.testing.expect(parsePPid("PPid:\tnotanumber\n") == null);
    // a truncated read must not answer from a half line
    try std.testing.expect(parsePPid("Name:\tsh\nPPi") == null);
}