The device manager supervises: hello, backoff, and the crash-loop cap (M18.1)

The manager is now a harness service on the well-known .device_manager
endpoint. Every driver spawns supervised; drivers with an assignment must
hello (device-manager-protocol, versioned) within a deadline enforced by
a timer sweep. Exit reasons drive the restart decision: clean exits stay
down, faults restart with 300/600/1200ms backoff, and three fast deaths
mark a driver failed instead of respawning forever. usb-xhci-bus is the
first conforming driver; the crash-test fixture claims a device, hellos,
and faults on purpose — each respawn re-proving claim release on death
through the manager's own path. maximum_tasks grows 16 -> 32: the
initial-ramdisk sweep (15 binaries at once) was intermittently
overflowing the static pool.

system/abi.zig

@@ -176,6 +176,7 @@ pub const ServiceId = enum(u32) {
     vfs = 1,
     input = 2,
     ps2_bus = 3, // the 8042 owner; child device drivers attach here for raw bytes
+    device_manager = 4, // the tree, the matcher, the supervisor (docs/device-manager.md)
     _,
 };
 

system/drivers/usb-xhci-bus/usb-xhci-bus.zig

@@ -1,58 +1,59 @@
 //! /system/drivers/usb-xhci-bus — the xHCI (USB 3) host-controller bus driver.
-//! The device manager spawns **one instance per controller** it discovers (a machine
-//! can carry several), passing the controller's device-tree id as argv[1]; this
-//! instance claims that device and no other, so multiple instances never fight over
-//! hardware. This increment proves the plumbing: parse the id, claim the controller,
-//! and report its MMIO window. The next increments map the registers and bring the
-//! controller up (reset, rings, port scan), then enumerate the USB devices on the
-//! bus with the usb-abi request builders and publish each with `device_register`.
+//! The device manager spawns **one instance per controller** it discovers (a
+//! machine can carry several), passing the controller's device-tree id as
+//! argv[1]; this instance claims that device and no other, so multiple
+//! instances never fight over hardware.
+//!
+//! M18.1 (this increment): a harness service and the first conforming driver of
+//! the device-manager protocol — claim the controller, `hello` the manager
+//! (role, version, assignment) inside its deadline, then serve. Controller
+//! bring-up (map the MMIO window, reset, port scan) and tree reports
+//! (`child_added` for each connected port) land in M18.2.
 
 const std = @import("std");
 const runtime = @import("runtime");
+const protocol = runtime.device_manager_protocol;
 const device = runtime.device;
 
-/// Format one whole log line and emit it in a single `debug_write`, so concurrent
-/// instances (one per controller) can never interleave mid-line.
+/// Format one whole log line and emit it in a single `debug_write`, so
+/// concurrent instances (one per controller) can never interleave mid-line.
 fn writeLine(comptime fmt: []const u8, arguments: anytype) void {
     var line: [128]u8 = undefined;
     _ = runtime.system.write(std.fmt.bufPrint(&line, fmt, arguments) catch return);
 }
 
-pub fn main(init: runtime.process.Init) void {
-    const argument = init.arguments.get(1) orelse {
-        _ = runtime.system.write("usb-xhci-bus: missing controller device id (argv[1])\n");
-        return;
-    };
-    const controller_id = std.fmt.parseInt(u64, argument, 10) catch {
-        writeLine("usb-xhci-bus: malformed controller device id '{s}'\n", .{argument});
-        return;
-    };
+var controller_id: u64 = protocol.no_device;
 
+/// Claim the assigned controller, find its register window, and hello the
+/// manager. Any failure returns false: the process exits cleanly, which the
+/// manager reads as "meant to stop" — a missing assignment is not a crash loop.
+fn initialise(endpoint: runtime.ipc.Handle) bool {
+    _ = endpoint;
     if (!device.claim(controller_id)) {
         writeLine("usb-xhci-bus: unable to claim controller device {d}\n", .{controller_id});
-        return;
+        return false;
     }
 
     // Fetch our own descriptor back for the controller's resources.
     const buffer = runtime.allocator().alloc(device.DeviceDescriptor, 64) catch {
         _ = runtime.system.write("usb-xhci-bus: out of memory\n");
-        return;
+        return false;
     };
     const total = device.enumerate(buffer);
     const descriptor = for (buffer[0..@min(total, buffer.len)]) |d| {
         if (d.id == controller_id) break d;
     } else {
         writeLine("usb-xhci-bus: device {d} not in the device tree\n", .{controller_id});
-        return;
+        return false;
     };
 
-    // The controller's operational registers live behind BAR0, enumerated as the
-    // device's first memory resource.
+    // The controller's operational registers live behind BAR0, enumerated as
+    // the device's first memory resource.
     const register_window = for (descriptor.resources[0..@intCast(descriptor.resource_count)]) |resource| {
         if (resource.kind == @intFromEnum(device.ResourceKind.memory)) break resource;
     } else {
         writeLine("usb-xhci-bus: controller device {d} has no MMIO window\n", .{controller_id});
-        return;
+        return false;
     };
     writeLine("usb-xhci-bus: claimed controller device {d} (registers at 0x{x}, {d} bytes)\n", .{
         controller_id,
@@ -60,9 +61,53 @@ pub fn main(init: runtime.process.Init) void {
         register_window.len,
     });
 
-    // Controller bring-up (map the window, reset, rings, port scan) is the next
-    // increment; stay resident as the bus's supervisor in the meantime.
-    while (true) runtime.system.sleep(1000);
+    // The handshake: role, protocol version, assignment — inside the manager's
+    // deadline (the lookup retries cover the manager still registering).
+    var manager: ?runtime.ipc.Handle = null;
+    var tries: u32 = 0;
+    while (manager == null and tries < 100) : (tries += 1) {
+        manager = runtime.ipc.lookup(.device_manager);
+        if (manager == null) runtime.system.sleep(20);
+    }
+    const h = manager orelse {
+        _ = runtime.system.write("usb-xhci-bus: no device manager to hello\n");
+        return false;
+    };
+    const hello = protocol.Hello{ .role = @intFromEnum(protocol.Role.bus), .device_id = controller_id };
+    var reply: [protocol.message_maximum]u8 = undefined;
+    const n = runtime.ipc.call(h, std.mem.asBytes(&hello), &reply) catch {
+        _ = runtime.system.write("usb-xhci-bus: hello call failed\n");
+        return false;
+    };
+    if (n < protocol.reply_size or std.mem.bytesToValue(protocol.HelloReply, reply[0..protocol.reply_size]).status != 0) {
+        _ = runtime.system.write("usb-xhci-bus: hello refused\n");
+        return false;
+    }
+    _ = runtime.system.write("usb-xhci-bus: hello acknowledged\n");
+    return true;
+}
+
+/// No bus protocol to serve yet — transfer requests arrive with the USB track.
+fn onMessage(message: []const u8, reply: []u8, sender: u32) usize {
+    _ = message;
+    _ = reply;
+    _ = sender;
+    return 0;
+}
+
+pub fn main(init: runtime.process.Init) void {
+    const argument = init.arguments.get(1) orelse {
+        _ = runtime.system.write("usb-xhci-bus: missing controller device id (argv[1])\n");
+        return;
+    };
+    controller_id = std.fmt.parseInt(u64, argument, 10) catch {
+        writeLine("usb-xhci-bus: malformed controller device id '{s}'\n", .{argument});
+        return;
+    };
+    runtime.service.run(protocol.message_maximum, .{
+        .init = initialise,
+        .on_message = onMessage,
+    });
 }
 
 pub const panic = runtime.panic;

system/kernel/tests.zig

@@ -138,6 +138,8 @@ pub fn run(case: []const u8, boot_information: *const BootInformation) void {
         vfsClientDeathTest(boot_information);
     } else if (eql(case, "signals")) {
         signalsTest(boot_information);
+    } else if (eql(case, "driver-restart")) {
+        driverRestartTest(boot_information);
     } else if (eql(case, "initial-ramdisk")) {
         initialRamdiskTest(boot_information);
     } else if (eql(case, "vfs")) {
@@ -1654,6 +1656,44 @@ fn signalsTest(boot_information: *const BootInformation) void {
     result();
 }
 
+/// M18.1: the device manager's restart machinery, end to end. In test-restart
+/// mode the manager also supervises crash-test: a fixture that claims device 0,
+/// hellos, and faults. The scenario asserts three markers in order — the real
+/// xHCI driver hellos clean and stays; crash-test is restarted with backoff
+/// (each respawn re-claiming the device the dead instance held, M17.1 through
+/// the manager's path); the crash loop caps and the manager gives up.
+fn driverRestartTest(boot_information: *const BootInformation) void {
+    log("DANOS-TEST-BEGIN: driver-restart\n", .{});
+    if (boot_information.initial_ramdisk_len == 0) {
+        check("bootloader handed over an initial_ramdisk", false);
+        result();
+        return;
+    }
+    const image = @as([*]const u8, @ptrFromInt(boot_handoff.physicalToVirtual(boot_information.initial_ramdisk_base)))[0..boot_information.initial_ramdisk_len];
+    const rd = initial_ramdisk.Reader.init(image) orelse {
+        check("initial_ramdisk image is valid", false);
+        result();
+        return;
+    };
+
+    process.setInitialRamdisk(image); // the manager system_spawns drivers by name
+    process.write_count = 0;
+    var manager: u32 = 0;
+    var i: u32 = 0;
+    while (i < rd.count) : (i += 1) {
+        const item = rd.entry(i) orelse continue;
+        if (!eql(item.name, "device-manager")) continue;
+        manager = process.spawnProcessSupervised(item.blob, 4, &.{ "device-manager", "test-restart" }, scheduler.currentId(), null) catch 0;
+        break;
+    }
+    check("device-manager spawned in test-restart mode", manager != 0);
+    // The assertions live in the harness: its expect regex requires, in order,
+    // the xHCI hello ack, a crash-test restart, and the crash-loop cap — read
+    // from the whole serial capture, immune to the transient-line races a
+    // write_buffer poll would have here (many processes log concurrently).
+    result();
+}
+
 /// The whole user-side surface at once: spawn process-test's supervisor role,
 /// which — entirely from ring 3 — creates an exit endpoint, spawns its two
 /// children supervised, sees them in process_enumerate, kills them (one blocked,

system/parameters.zig

@@ -16,8 +16,11 @@
 pub const maximum_cpus = 128;
 
 /// Maximum tasks (kernel threads) alive at once — the static task-table size. Each
-/// online core consumes one slot for its idle task, plus task 0 on the BSP.
-pub const maximum_tasks = 16;
+/// online core consumes one slot for its idle task, plus task 0 on the BSP. Sized
+/// for the initial-ramdisk sweep (15 bundled binaries spawned at once) plus the
+/// device manager's supervised children with room to grow — at 16 the sweep
+/// started failing spawns once the bundle passed a dozen binaries.
+pub const maximum_tasks = 32;
 
 /// Each task's kernel stack (also each AP's bring-up stack), in bytes.
 pub const kernel_stack_size = 16 * 1024;

system/services/crash-test/crash-test.zig

@@ -0,0 +1,44 @@
+//! crash-test — a test fixture, not a driver: claims the device it is assigned,
+//! hellos the device manager, announces itself, then faults on purpose. The
+//! driver-restart scenario drives the manager's whole restart machinery with
+//! it: fault → exit reason → backoff → respawn → the **same claim succeeding
+//! again** (claim release on death, M17.1, through the manager's path) → the
+//! crash-loop cap. Spawned bare (the initial-ramdisk sweep starts every bundled
+//! binary), it exits silently so it cannot derange other tests.
+
+const std = @import("std");
+const runtime = @import("runtime");
+const protocol = runtime.device_manager_protocol;
+
+pub fn main(init: runtime.process.Init) void {
+    const argument = init.arguments.get(1) orelse return; // bare: stay silent
+    const assigned = std.fmt.parseInt(u64, argument, 10) catch return;
+
+    // The respawn only reaches this line because the kernel released the
+    // previous instance's claim at death. A failed claim exits cleanly — the
+    // manager reads "meant to stop" and the scenario fails loudly by silence.
+    if (!runtime.device.claim(assigned)) {
+        _ = runtime.system.write("crash-test: claim failed\n");
+        return;
+    }
+
+    var manager: ?runtime.ipc.Handle = null;
+    var tries: u32 = 0;
+    while (manager == null and tries < 100) : (tries += 1) {
+        manager = runtime.ipc.lookup(.device_manager);
+        if (manager == null) runtime.system.sleep(20);
+    }
+    const h = manager orelse return;
+    const hello = protocol.Hello{ .role = @intFromEnum(protocol.Role.device), .device_id = assigned };
+    var reply: [protocol.message_maximum]u8 = undefined;
+    _ = runtime.ipc.call(h, std.mem.asBytes(&hello), &reply) catch return;
+
+    _ = runtime.system.write("crash-test: faulting now\n");
+    const poison: *volatile u32 = @ptrFromInt(0xdead0000);
+    poison.* = 1; // the restart machinery's fuel: a real segmentation fault
+}
+
+pub const panic = runtime.panic;
+comptime {
+    _ = &runtime.start._start; // pull the runtime entry shim into the image
+}

system/services/device-manager/device-manager-protocol.zig

@@ -0,0 +1,58 @@
+//! The device-manager protocol (docs/device-manager.md): what drivers and
+//! applications say to the device manager over its well-known endpoint. The
+//! vfs-protocol pattern — extern-struct messages, a version in the handshake,
+//! reserved fields — so both sides depend on the contract by name. Deliberately
+//! contains nothing lifecycle-shaped: stopping, liveness (the zero-length ping),
+//! and exit reasons are the universal vocabulary of
+//! docs/process-lifecycle.md, not this protocol.
+
+/// The protocol version a driver states in its hello. A manager that cannot
+/// serve a driver's version refuses the hello, and the mismatch is loud at
+/// startup instead of quiet corruption later.
+pub const version: u16 = 1;
+
+/// What kind of driver is talking (docs/driver-model.md's shapes).
+pub const Role = enum(u8) {
+    /// Owns a controller and reports the devices behind it (`child_added`).
+    bus = 1,
+    /// Serves one device, reached through a bus's transfer protocol.
+    device = 2,
+};
+
+/// The message kinds. `child_added`/`child_removed` land in M18.2;
+/// `enumerate`/`subscribe` in M18.3.
+pub const Operation = enum(u8) {
+    hello = 1,
+};
+
+/// `Hello.device_id` for a driver that serves no enumerated device (a test
+/// fixture, a synthetic source).
+pub const no_device: u64 = ~@as(u64, 0);
+
+/// The handshake, sent once by every driver the manager spawns — the manager's
+/// one self-enforced deadline: spawned and silent past it means wrong binary,
+/// wrong version, or wedged before main, and the stop sequence follows.
+pub const Hello = extern struct {
+    operation: u8 = @intFromEnum(Operation.hello),
+    /// A Role value.
+    role: u8,
+    /// The protocol version this driver was built against (`version`).
+    version: u16 = version,
+    reserved: u32 = 0,
+    /// The device this driver was assigned (its argv[1]), or `no_device`.
+    device_id: u64,
+};
+
+pub const hello_size = @sizeOf(Hello);
+
+/// The manager's answer to a hello. Nonzero status = refused (version mismatch,
+/// unknown sender); a refused driver should exit cleanly.
+pub const HelloReply = extern struct {
+    status: i32,
+    reserved: u32 = 0,
+};
+
+pub const reply_size = @sizeOf(HelloReply);
+
+/// Upper bound on any message in this protocol — sizes the endpoint buffers.
+pub const message_maximum = 64;

system/services/device-manager/device-manager.zig

@@ -1,20 +1,24 @@
 //! /system/services/device-manager — the ring-3 process that turns the device
-//! tree into a running system. The kernel enumerates the hardware and enforces the
-//! claim capability (mechanism); this decides *which driver serves which device*
-//! and, eventually, spawns it (policy). Keeping that split in user space is the
-//! whole point of the microkernel: the manager is an ordinary, restartable process
-//! with no special privilege — it uses the same `device_*` system calls any process
-//! could ([drivers.md](../../../docs/drivers.md), [driver-model.md]).
+//! tree into a running system: **the matcher and the supervisor**
+//! (docs/device-manager.md). The kernel enumerates the hardware and enforces the
+//! claim capability (mechanism); this decides which driver serves which device,
+//! spawns it, and keeps it alive (policy). Keeping that split in user space is
+//! the whole point of the microkernel: the manager is an ordinary, restartable
+//! process with no special privilege.
 //!
-//! Increment 2 (this file): enumerate /system/devices, *match* each device to a
-//! driver, and *spawn* it with `system_spawn` — the kernel loads the named binary
-//! from the initial-ramdisk as a fresh ring-3 process. On QEMU this discovers the
-//! HPET, decides `hpet` serves it, and brings that driver all the way up. (The
-//! kernel still auto-spawns the whole initial-ramdisk at boot; increment 3 removes
-//! that redundancy so the manager is the sole owner of driver spawning.)
+//! M18.1 (this increment): the manager is a harness service on the well-known
+//! `.device_manager` endpoint. Every driver is spawned **supervised** — exit
+//! notifications land in the same loop as protocol messages. Drivers with an
+//! assignment must `hello` within a deadline or be stopped; a driver that dies
+//! is restarted with backoff, and a crash loop (three fast deaths) marks it
+//! failed instead of respawning forever. Exit reasons (M17.2) drive the
+//! decision: a clean exit meant to stop; only faults and missed deadlines
+//! restart. Tree reports (`child_added`) land in M18.2.
+
 const std = @import("std");
 const runtime = @import("runtime");
 const acpi_ids = @import("acpi-ids");
+const protocol = runtime.device_manager_protocol;
 const device = runtime.device;
 const system = runtime.system;
 
@@ -27,9 +31,8 @@ fn writeLine(comptime fmt: []const u8, arguments: anytype) void {
 }
 
 /// The driver that serves each device — the policy table. In a fuller system
-/// this comes from the drivers describing what they bind (or a manifest under
-/// /system/drivers); for now it is a small static map, which is enough to prove the
-/// manager reads the tree and decides. `null` = no driver for this class yet.
+/// this comes from a manifest (docs/device-manager.md: the third bus type
+/// triggers it); for now a static map. `null` = no driver for this class yet.
 fn driverFor(d: device.DeviceDescriptor) ?[]const u8 {
     // detect device via DeviceClass
     if (d.class == @intFromEnum(device.DeviceClass.timer)) return "hpet";
@@ -47,10 +50,9 @@ fn driverFor(d: device.DeviceDescriptor) ?[]const u8 {
 /// pci-class.zig decodes.
 const xhci_pci_class: u64 = 0x0C_03_30;
 
-/// The bus driver that serves a PCI function, or null. Unlike the singleton drivers
-/// in `driverFor`, a machine can carry several identical controllers — so the caller
-/// spawns one driver instance *per device*, passing the device id as argv[1] for the
-/// instance to claim.
+/// The bus driver that serves a PCI function, or null. A machine can carry
+/// several identical controllers — one driver instance per device, the id as
+/// argv[1]. These drivers speak the protocol: a hello is expected.
 fn pciDriverFor(d: device.DeviceDescriptor) ?[]const u8 {
     if (d.class != @intFromEnum(device.DeviceClass.pci_device)) return null;
     return switch (d.pci_class) {
@@ -59,24 +61,169 @@ fn pciDriverFor(d: device.DeviceDescriptor) ?[]const u8 {
     };
 }
 
-/// Spawn one instance of `driver_name` to serve the specific device `id` — the id
-/// arrives as argv[1]. No isProcessRunning gate here: the name alone cannot tell two
-/// instances apart, and this manager is the sole spawner of drivers.
-fn spawnForDevice(driver_name: []const u8, id: u64) void {
-    var text: [20]u8 = undefined;
-    const id_text = std.fmt.bufPrint(&text, "{d}", .{id}) catch return;
-    if (system.spawnWithArguments(driver_name, &.{id_text}) != null) {
-        writeLine("device-manager: spawned {s} for device {d}\n", .{ driver_name, id });
+// --- supervision -------------------------------------------------------------
+
+/// How long a protocol driver has to hello after its spawn.
+const hello_deadline_ms: u64 = 3000;
+/// Deaths faster than this count toward the crash loop; slower ones reset it.
+const fast_death_ns: u64 = 2_000_000_000;
+/// Consecutive fast deaths before the manager gives up on a driver.
+const crash_loop_cap: u32 = 3;
+/// Restart backoff: base << (restarts - 1), so 300 ms, 600 ms, 1200 ms.
+const backoff_base_ms: u64 = 300;
+
+const DriverState = enum {
+    awaiting_hello, // spawned; the deadline is armed (protocol drivers only)
+    running,
+    restarting, // dead; respawn due at restart_due_ns
+    stopped, // exited cleanly — it meant to; not restarted
+    failed, // crash loop, or unspawnable; the manager gave up
+};
+
+const Driver = struct {
+    used: bool = false,
+    name_buffer: [24]u8 = undefined,
+    name_len: usize = 0,
+    // The assigned device id (becomes argv[1]), or protocol.no_device.
+    device_id: u64 = protocol.no_device,
+    // Whether this driver speaks the protocol (hello expected, deadline
+    // enforced). Legacy drivers (hpet, ps2-bus) are supervised and restarted
+    // but not yet required to hello.
+    speaks_protocol: bool = false,
+    process_id: u32 = 0,
+    state: DriverState = .running,
+    restarts: u32 = 0,
+    spawn_ns: u64 = 0,
+    hello_deadline_ns: u64 = 0,
+    restart_due_ns: u64 = 0,
+
+    fn name(driver: *const Driver) []const u8 {
+        return driver.name_buffer[0..driver.name_len];
+    }
+};
+
+const maximum_drivers = 16;
+var drivers: [maximum_drivers]Driver = .{Driver{}} ** maximum_drivers;
+var manager_endpoint: runtime.ipc.Handle = 0;
+var test_restart_mode = false;
+
+fn driverByProcess(process_id: u32) ?*Driver {
+    for (&drivers) |*driver| {
+        if (driver.used and driver.process_id == process_id) return driver;
+    }
+    return null;
+}
+
+/// Whether a singleton driver is already in the table (two ACPI nodes can both
+/// map to ps2-bus; one instance serves both).
+fn alreadySupervised(name: []const u8) bool {
+    for (&drivers) |*driver| {
+        if (driver.used and std.mem.eql(u8, driver.name(), name)) return true;
+    }
+    return false;
+}
+
+/// Record a driver in the table and spawn its first instance.
+fn addDriver(name: []const u8, device_id: u64, speaks_protocol: bool) void {
+    for (&drivers) |*driver| {
+        if (driver.used) continue;
+        const n = @min(name.len, driver.name_buffer.len);
+        @memcpy(driver.name_buffer[0..n], name[0..n]);
+        driver.name_len = n;
+        driver.device_id = device_id;
+        driver.speaks_protocol = speaks_protocol;
+        driver.used = true;
+        spawnDriver(driver);
+        return;
+    }
+    writeLine("device-manager: driver table full; cannot supervise {s}\n", .{name});
+}
+
+/// (Re)spawn a driver instance: supervised on the manager's own endpoint, the
+/// device id as argv[1] when it has one, the hello deadline armed when it
+/// speaks the protocol.
+fn spawnDriver(driver: *Driver) void {
+    var id_text: [20]u8 = undefined;
+    var arguments: [1][]const u8 = undefined;
+    var argument_count: usize = 0;
+    if (driver.device_id != protocol.no_device) {
+        arguments[0] = std.fmt.bufPrint(&id_text, "{d}", .{driver.device_id}) catch return;
+        argument_count = 1;
+    }
+    const child = system.spawnSupervised(driver.name(), arguments[0..argument_count], manager_endpoint) orelse {
+        writeLine("device-manager: failed to spawn {s}\n", .{driver.name()});
+        driver.state = .failed;
+        return;
+    };
+    driver.process_id = child;
+    driver.spawn_ns = system.clock();
+    if (driver.speaks_protocol) {
+        driver.state = .awaiting_hello;
+        driver.hello_deadline_ns = driver.spawn_ns + hello_deadline_ms * 1_000_000;
+        _ = system.timerOnce(manager_endpoint, hello_deadline_ms + 100);
     } else {
-        writeLine("device-manager: failed to spawn {s} for device {d}\n", .{ driver_name, id });
+        driver.state = .running;
+    }
+    if (driver.device_id != protocol.no_device) {
+        writeLine("device-manager: spawned {s} for device {d}\n", .{ driver.name(), driver.device_id });
+    } else {
+        writeLine("device-manager: spawned {s}\n", .{driver.name()});
     }
 }
 
-pub fn main() void {
+/// A driver died. The exit reason (M17.2) is the whole decision: a clean exit
+/// meant to stop; anything else restarts with backoff until the crash-loop cap.
+fn onDriverExit(driver: *Driver) void {
+    const reason = runtime.process.exitReason(driver.process_id) orelse .fault;
+    if (reason == .exited) {
+        driver.state = .stopped;
+        writeLine("device-manager: {s} exited cleanly; not restarting\n", .{driver.name()});
+        return;
+    }
+    const now = system.clock();
+    const alive_ns = now - driver.spawn_ns;
+    driver.restarts = if (alive_ns < fast_death_ns) driver.restarts + 1 else 1;
+    if (driver.restarts >= crash_loop_cap) {
+        driver.state = .failed;
+        writeLine("device-manager: {s} is failing repeatedly (crash loop); giving up\n", .{driver.name()});
+        return;
+    }
+    const delay_ms = backoff_base_ms << @intCast(driver.restarts - 1);
+    driver.state = .restarting;
+    driver.restart_due_ns = now + delay_ms * 1_000_000;
+    writeLine("device-manager: restarting {s} in {d} ms (died: {s})\n", .{ driver.name(), delay_ms, @tagName(reason) });
+    _ = system.timerOnce(manager_endpoint, delay_ms + 50);
+}
+
+/// A timer landed: sweep every deadline. Overdue hellos are killed (the exit
+/// notification then routes through the normal restart policy); due restarts
+/// respawn. Timers carry no id on purpose — the table is the state, and one
+/// sweep serves every armed deadline.
+fn sweepDeadlines() void {
+    const now = system.clock();
+    for (&drivers) |*driver| {
+        if (!driver.used) continue;
+        switch (driver.state) {
+            .awaiting_hello => if (now >= driver.hello_deadline_ns) {
+                writeLine("device-manager: {s} missed its hello deadline\n", .{driver.name()});
+                _ = system.kill(driver.process_id);
+                // The exit notification finishes the job via onDriverExit.
+            },
+            .restarting => if (now >= driver.restart_due_ns) spawnDriver(driver),
+            else => {},
+        }
+    }
+}
+
+// --- the harness callbacks -----------------------------------------------------
+
+fn initialise(endpoint: runtime.ipc.Handle) bool {
+    manager_endpoint = endpoint;
+
     // Enumerate into a heap buffer (too big for the one-page user stack).
     const buffer = runtime.allocator().alloc(device.DeviceDescriptor, 64) catch {
         _ = runtime.system.write("device-manager: out of memory\n");
-        return;
+        return false;
     };
     const total = device.enumerate(buffer);
     const count = @min(total, buffer.len);
@@ -85,28 +232,74 @@ pub fn main() void {
     for (buffer[0..count]) |descriptor| {
         if (pciDriverFor(descriptor)) |driver_name| {
             matched += 1;
-            spawnForDevice(driver_name, descriptor.id);
+            addDriver(driver_name, descriptor.id, true);
             continue;
         }
         const driver_name = driverFor(descriptor) orelse continue;
         matched += 1;
-        if (!system.isProcessRunning(driver_name)) {
-            if (runtime.system.spawn(driver_name) != null) {
-                writeLine("device-manager: spawned {s}\n", .{driver_name});
-            } else {
-                writeLine("device-manager: failed to spawn {s}\n", .{driver_name});
-            }
-        } else {
-            writeLine("device-manager: already spawned {s}\n", .{driver_name});
+        // Skip a singleton that is already alive (the initial-ramdisk sweep test
+        // starts every bundled binary bare, this manager included) — spawning a
+        // second instance would only lose the claim race and churn the log.
+        if (!alreadySupervised(driver_name) and !system.isProcessRunning(driver_name)) {
+            addDriver(driver_name, protocol.no_device, false);
         }
     }
 
+    if (test_restart_mode) {
+        // The driver-restart scenario's fixture: claims device 0 (the tree
+        // root, otherwise unclaimed), hellos, then faults — driving backoff,
+        // re-claim-after-death, and the crash-loop cap deterministically.
+        addDriver("crash-test", 0, true);
+    }
+
     if (matched == 0) {
         _ = runtime.system.write("device-manager: no matchable devices\n");
+    } else {
+        _ = runtime.system.write("device-manager: ok\n");
+    }
+    return true;
+}
+
+fn onMessage(message: []const u8, reply: []u8, sender: u32) usize {
+    if (message.len < protocol.hello_size) return 0;
+    const hello = std.mem.bytesToValue(protocol.Hello, message[0..protocol.hello_size]);
+    if (hello.operation != @intFromEnum(protocol.Operation.hello)) return 0;
+
+    var status: i32 = 0;
+    if (hello.version != protocol.version) {
+        status = -1;
+        writeLine("device-manager: refused hello (version {d}) from process {d}\n", .{ hello.version, sender });
+    } else if (driverByProcess(sender)) |driver| {
+        driver.state = .running;
+        writeLine("device-manager: hello from {s} (device {d})\n", .{ driver.name(), hello.device_id });
+    } else {
+        status = -1;
+        writeLine("device-manager: hello from unknown process {d}\n", .{sender});
+    }
+    const hello_reply = protocol.HelloReply{ .status = status };
+    @memcpy(reply[0..protocol.reply_size], std.mem.asBytes(&hello_reply));
+    return protocol.reply_size;
+}
+
+fn onNotification(badge: u64) void {
+    if (badge & runtime.ipc.notify_exit_bit != 0) {
+        const dead: u32 = @intCast(badge & ~(runtime.ipc.notify_badge_bit | runtime.ipc.notify_exit_bit));
+        if (driverByProcess(dead)) |driver| onDriverExit(driver);
         return;
     }
-    _ = runtime.system.write("device-manager: ok\n");
-    while (true) runtime.system.sleep(1000);
+    if (badge & runtime.ipc.notify_timer_bit != 0) sweepDeadlines();
+}
+
+pub fn main(init: runtime.process.Init) void {
+    if (init.arguments.get(1)) |mode| {
+        test_restart_mode = std.mem.eql(u8, mode, "test-restart");
+    }
+    runtime.service.run(protocol.message_maximum, .{
+        .service = .device_manager,
+        .init = initialise,
+        .on_message = onMessage,
+        .on_notification = onNotification,
+    });
 }
 
 pub const panic = runtime.panic;