diff --git a/bend2/comp.ts b/bend2/comp.ts
index ff0ca30fb..f5feda076 100644
--- a/bend2/comp.ts
+++ b/bend2/comp.ts
@@ -1304,9 +1304,13 @@ export function io_type(book: Bend.Book): HTerm | null {
   return xs?.length === 1 ? xs[0] : null;
 }
 
+function io_status(): boolean {
+  return ty_adt(io_type(FL.book))?.k === "U32";
+}
+
 export function io_run(book: Bend.Book, args: string[]): number {
   const src = `${js_lib(book)}\n${RUNTIME_MAIN}\ncli_args = ${
-    JSON.stringify(args)};\nreturn io_run(${js_sat("main")});`;
+    JSON.stringify(args)};\nreturn io_run(${js_sat("main")}, ${io_status()});`;
   return new Function("require", src)(import.meta.require);
 }
 
@@ -2883,6 +2887,7 @@ export function compile_book(book: Bend.Book): string {
   `#define WL_TABLE ${entries.map((s) => `WL_X(${s.fid})`).join(" ")
     } WL_X(FID_EXIT)`, `#define MAIN_FID ${seg_fid("main")}`,
   `#define MAIN_PURE ${Number(show !== null)}`,
+  `#define MAIN_STATUS ${Number(io_status())}`,
   `#define BLK_SHR ${Number(FL.hot.has("t:Array"))}`);
   const tabs = [defs.join("\n"), ...[...FL.tabs].map(([r, i]) =>
     `CONSTV u64 TAB_${i}[] = { ${r} };`)].join("\n\n");
@@ -3247,7 +3252,7 @@ export function js_book(book: Bend.Book): string {
   const show = show_main();
   return `${lib}\n${RUNTIME_MAIN}\ncli(process.argv.slice(1));\nio_exit(${
     js_sat("main")}, ${JSON.stringify(show && show.map((c) =>
-      typeof c === "string" ? Bend.name_key(c) : c))});`;
+      typeof c === "string" ? Bend.name_key(c) : c))}, ${io_status()});`;
 }
 
 // RuntimeC
@@ -5372,6 +5377,13 @@ typedef Term (*Effect)(Env e, Term* f, IoWork* w);
 Effect io_eff_rows[sizeof CID_T / sizeof *CID_T];
 static u32    io_live;
 
+static int io_code(u32 code) {
+  if (code > 255) {
+    err_fail("exit status must be between 0 and 255");
+  }
+  return code;
+}
+
 static u64 io_tick(void) {
   struct timespec ts;
   clock_gettime(CLOCK_MONOTONIC, &ts);
@@ -5418,6 +5430,9 @@ static IoWork* io_runs;
 static IoWork* io_park;
 static IoWork* io_jobs;
 
+static IoWork* io_root;
+static u32     io_result;
+
 static void io_push(IoWork** q, IoWork* a) {
   IoWork* l = *q != NULL ? *q : a;
   a->next = l->next;
@@ -5440,9 +5455,7 @@ static void io_spawn(Term m) {
   io_live += 1;
 }
 
-// io_park stays in deadline order (time 0, none, sorts last; ties keep
-// their park order), so io_wait wakes due timers in the order they expire.
-// It links both ways, so io_park_cut drops a waiter in O(1).
+// Deadline order: time 0 last, stable ties. Doubly linked for O(1) removal.
 static void io_park_add(IoWork* w) {
   IoWork* p = io_park;
   if (p == NULL || p->time - 1 <= w->time - 1) {
@@ -5882,6 +5895,10 @@ static void io_step(Env e, IoWork* a) {
     Term req = corpus_eval(e.mem, term_tsk(FID(Clo~apply), ap));
     u32  c   = (u32)term_aux(req);
     if (c == CID(Emit)) {
+      if (MAIN_STATUS && a == io_root) {
+        io_result = e.mem[term_peek(e.mem, req)];
+        io_root = NULL;
+      }
       term_drop(e, req);
       free(a);
       io_live -= 1;
@@ -5891,8 +5908,9 @@ static void io_step(Env e, IoWork* a) {
     u32  n = cid_arity(c);
     spare_free(e, cls_fit(n), ctr_take(e, req, n, fs));
     if (c == CID(Halt)) {
+      int code = io_code(fs[0]);
       io_errs(e, fs[1]);
-      exit((int)(u32)fs[0]);
+      exit(code);
     }
     if (io_eff_rows[c] == NULL) {
       err_fail("an alien request");
@@ -5921,6 +5939,9 @@ OUTLINE void io_loop(u64* H) {
   return;
 #endif
   io_spawn(m);
+  if (MAIN_STATUS) {
+    io_root = io_runs;
+  }
   u64 look = 0;
   for (u32 n = 0;; n += 1) {
     if (io_runs == NULL) {
@@ -6012,7 +6033,7 @@ int main(int argc, char** argv) {
   }
   io_loop(corpus_setup(why == NULL, thr > 0 ? thr : cpu_count(), mem));
   io_sync();
-  return 0;
+  return MAIN_STATUS ? io_code(io_result) : 0;
 }
 
 #endif
@@ -6154,21 +6175,23 @@ function show_val(D, d, v, chain) {
 // Io
 // ==
 
-// Apple arm64 passes variadic fcntl flags on the stack, so io_sys
-// binds fcntl there with the flags as the ninth fixed argument. A
-// parked effect waits for fd (a write when out) or until at
-// (performance.now()), either one undefined when unused; once due, io_wait
-// calls more at once, as C calls pack, and resumes k with its value, while
-// undefined parks it again. The waits stay in deadline order, as io_park
-// does in C.
+// Apple arm64 fcntl flags are stack argument 9. Deadline-ordered waits accept
+// fd/out or at (performance.now()); more (C: pack) resumes k, undefined re-parks.
+
+function io_code(code) {
+  if (code < 0 || code > 255) {
+    throw "bend: exit status must be between 0 and 255";
+  }
+  return code;
+}
 
-function io_exit(main, show) {
+function io_exit(main, show, status) {
   try {
     if (show !== null) {
       io_out(1, io_bytes(show_val(show, 0, run_loop(main()), 0) + "\n"));
       process.exit(0);
     }
-    process.exit(io_run(main));
+    process.exit(io_run(main, status));
   } catch (e) {
     io_errs(String(e));
     process.exit(1);
@@ -6351,16 +6374,20 @@ function io_park_on(fd, out, k, more, at) {
   ws.splice(i + 1, 0, { fd, out, k, more, at });
 }
 
-function io_run(m) {
+function io_run(m, status) {
   const io = { runs: [], live: 0, waits: [] };
   globalThis.BEND_IO = io;
+  let code = 0;
   try {
-    io_push(run_loop(m()), (x) => ({ $: "Emit", value: x }), true);
+    io_push(run_loop(m()), (x) => {
+      if (status) code = x;
+      return { $: "Emit", value: x };
+    }, true);
     let look = 0;
     for (let n = 0;; n += 1) {
       if (io.runs.length === 0) {
         if (io.live === 0) {
-          return 0;
+          return status ? io_code(code) : 0;
         }
         if (io.waits.length === 0) {
           io_errs("bend: deadlock: every computation waits on a channel");
@@ -6384,8 +6411,9 @@ function io_run(m) {
           break;
         }
         if (op.$ === "Halt") {
+          const halt = io_code(op.code);
           io_errs(op.message);
-          return op.code;
+          return halt;
         }
         const run = $0eff[op.$];
         if (run === undefined) {
diff --git a/guide/GUIDE.md b/guide/GUIDE.md
index 87d0162f9..60fc535dd 100644
--- a/guide/GUIDE.md
+++ b/guide/GUIDE.md
@@ -368,13 +368,13 @@ def main() -> IO(Unit):
     IO.print(text)
 ```
 
-Every bind is annotated, and `x : T = v` binds a pure value in the middle of a
-block. A fallible effect answers `Result<&1, &1, U32 & String, A>`: `IO.try`
-unwraps it or exits with the error, and `IO.die` exits with your own. `IO.args`
-answers the command line, less the runtime's own options (a `--` ends them):
-its head is the program as invoked, like C's `argv[0]`. A handle (`File`,
-`Socket`, `Window`) is an affine, opaque value, so every effect on one hands it
-back beside its result, and no program can forge or reuse one.
+Binds need types; `x : T = v` binds a pure value. Fallible effects answer
+`Result<&1, &1, U32 & String, A>`; `IO.try` unwraps or exits with the error.
+`IO.die` exits with your message. An `IO(U32)` main exits silently with its
+result; `IO(Unit)` exits 0. Main/die codes outside 0..255 fail with a diagnostic
+and exit 1. `IO.args` includes the program name and omits runtime options before
+`--`. Handles are affine and opaque: effects return them beside results,
+preventing forging or reuse.
 
 `TCP.listen(host, port)` and `UDP.bind(host, port)` bind the IPv4 literal
 `host`: `"127.0.0.1"` serves this machine only, `"0.0.0.0"` every interface.
diff --git a/tests/io/channel_fork_join.bend b/tests/io/channel_fork_join.bend
index 69fdf48c5..13e6d7fba 100644
--- a/tests/io/channel_fork_join.bend
+++ b/tests/io/channel_fork_join.bend
@@ -1,9 +1,7 @@
 # io (io_channel_fork_join): bend3's channel round trip on bend's
 # surface. A spawned producer sends 11 and 22 into a channel with
 # room for four and closes it; main receives both in order (Some) and
-# prints 1122, then forks a pure value and joins it. main's value is
-# not the exit status: the process exits 0 once every computation
-# has answered.
+# prints 1122, then forks and joins a pure value, returning success.
 import Base
 
 law producer:
@@ -40,7 +38,7 @@ def main.go(c):
     z : Unit <- IO.print(U32.show((got(x) * 100 + got(y) : U32)))
     t : Chan(U32) <- IO.fork(U32, IO.pure(U32, 5))
     w : U32 <- IO.join(U32, t)
-    return w
+    return 0
 
 law main:
   IO(U32)
diff --git a/tests/io/die_status_invalid.bend b/tests/io/die_status_invalid.bend
new file mode 100644
index 000000000..9a673101c
--- /dev/null
+++ b/tests/io/die_status_invalid.bend
@@ -0,0 +1,7 @@
+import Base
+
+def main() -> IO(Unit):
+  IO.die(Unit, 256, "x")
+
+#|bend: exit status must be between 0 and 255
+#|exit 1
diff --git a/tests/io/exit_status.bend b/tests/io/exit_status.bend
new file mode 100644
index 000000000..21b710473
--- /dev/null
+++ b/tests/io/exit_status.bend
@@ -0,0 +1,9 @@
+import Base
+
+def main() -> IO(U32):
+  do IO<U32>:
+    IO.print_err("usage error")
+    return 2
+
+#|usage error
+#|exit 2
diff --git a/tests/io/exit_status_children.bend b/tests/io/exit_status_children.bend
new file mode 100644
index 000000000..db9e4613b
--- /dev/null
+++ b/tests/io/exit_status_children.bend
@@ -0,0 +1,22 @@
+import Base
+
+def grandchild() -> IO(U32):
+  do IO<U32>:
+    IO.print("grandchild finished")
+    return 9
+
+def child() -> IO(U32):
+  do IO<U32>:
+    IO.sleep(1)
+    IO.spawn(U32, grandchild())
+    IO.print("child finished")
+    return 7
+
+def main() -> IO(U32):
+  do IO<U32>:
+    IO.spawn(U32, child())
+    return 2
+
+#|child finished
+#|grandchild finished
+#|exit 2
diff --git a/tests/io/exit_status_invalid.bend b/tests/io/exit_status_invalid.bend
new file mode 100644
index 000000000..2b8880caa
--- /dev/null
+++ b/tests/io/exit_status_invalid.bend
@@ -0,0 +1,8 @@
+import Base
+
+def main() -> IO(U32):
+  do IO<U32>:
+    return 256
+
+#|bend: exit status must be between 0 and 255
+#|exit 1
diff --git a/tests/io/exit_status_limit.bend b/tests/io/exit_status_limit.bend
new file mode 100644
index 000000000..d6d84a29b
--- /dev/null
+++ b/tests/io/exit_status_limit.bend
@@ -0,0 +1,7 @@
+import Base
+
+def main() -> IO(U32):
+  do IO<U32>:
+    return 255
+
+#|exit 255
diff --git a/tests/io/fiber_fork_join.bend b/tests/io/fiber_fork_join.bend
index 8b90b547e..0295a076e 100644
--- a/tests/io/fiber_fork_join.bend
+++ b/tests/io/fiber_fork_join.bend
@@ -12,6 +12,6 @@ def main():
     a : Unit <- IO.join(Unit, t)
     u : Unit <- IO.spawn(Unit, IO.pure(Unit, Unit{}))
     b : Unit <- IO.print("joined")
-    return 9
+    return 0
 
 #|joined
diff --git a/tests/io/print_write.bend b/tests/io/print_write.bend
index 5a7c64b8e..79b7beb94 100644
--- a/tests/io/print_write.bend
+++ b/tests/io/print_write.bend
@@ -1,8 +1,6 @@
 # io (io_print_write): print (newline) and write (raw) sequence in
 # order through the event loop, and U32.show feeds a computed string
-# to an effect. Ported from bend3; bend3's result-word exit code was
-# bend3 CLI behavior -- bend's loop discards the Emit word, and the
-# interp lane pins no exit line.
+# to an effect.
 import Base
 
 law main:
@@ -13,7 +11,7 @@ def main():
     a : Unit <- IO.print("hello")
     b : Unit <- IO.write("num=")
     c : Unit <- IO.print(U32.show((40 + 2 : U32)))
-    return 7
+    return 0
 
 #|hello
 #|num=42