fix: install a default guest kernel in setup, stop the runtime for the updater

CHANGELOG.md

@@ -6,6 +6,18 @@ All notable changes to this project are documented here. The format follows
 
 ## [Unreleased]
 
+### Fixed
+
+- `setup` now installs apple/container's recommended guest kernel when no default is configured.
+  apple/container cannot boot any VM without one, so on a fresh install — or one whose user data
+  was deleted, which `cleanup --all -d` does by design — every sandbox create failed at image
+  unpack with `default kernel not configured for architecture arm64`. The step is skipped when a
+  kernel is already set, and a failed download only warns (with the retry command) so `setup`
+  still repairs the rest of the wiring.
+- `update --all` now stops the container runtime before running apple/container's updater and
+  restarts it afterwards. The updater refuses to run while the runtime is up, so this step
+  previously always failed with "`container` is still running".
+
 ## [0.2.5] - 2026-08-02
 
 ### Fixed

README.md

@@ -157,6 +157,7 @@ source of truth. See [docs/architecture.md](docs/architecture.md).
 | anything looks broken | re-run `openshell-driver-applecontainer setup` — it repairs all wiring |
 | `openshell status` fails | `brew services info openshell`; gateway log: `/opt/homebrew/var/log/openshell/openshell-gateway.err.log` |
 | sandbox stuck / failed | driver log: `~/Library/Logs/openshell-driver-applecontainer.log`; failed sandboxes carry the guest console tail in their status |
+| `default kernel not configured for architecture` | apple/container has no guest kernel (a fresh install, or one whose data was deleted). `setup` installs one automatically; if its download failed, retry `container system kernel set --recommended` |
 | slow first create | the base image (~2.6 GB) is pulling; `setup` without `--no-pull` pre-pulls it |
 
 ## Configuration reference

cmd/openshell-driver-applecontainer/update.go

@@ -225,10 +225,20 @@ func updatePrerequisites(log *slog.Logger) {
 	}
 	const acUpdater = "/usr/local/bin/update-container.sh"
 	if _, err := os.Stat(acUpdater); err == nil {
+		// The updater refuses to run while the runtime is up ("`container` is
+		// still running"), so stop it first — same as the uninstall path does.
+		if err := streamCmd("container", "system", "stop"); err != nil {
+			log.Debug("container system stop", "err", err)
+		}
 		log.Info("update: updating apple/container (its updater needs sudo)")
 		if err := streamCmd(acUpdater); err != nil {
 			log.Warn("apple/container updater failed", "err", err)
 		}
+		// Bring the runtime back up; the driver needs it. (setup would also
+		// start it, but --no-setup must not leave it stopped.)
+		if err := streamCmd("container", "system", "start"); err != nil {
+			log.Warn("could not restart the container runtime; run `container system start`", "err", err)
+		}
 	}
 }
 

internal/hostsetup/cleanup_test.go

@@ -176,6 +176,40 @@ func TestCleanupAllDeleteDataPassesDeleteFlag(t *testing.T) {
 	}
 }
 
+// A fresh apple/container install (or one whose user data was deleted by
+// `cleanup --all -d`) has no default kernel, and then every sandbox create
+// fails at image unpack. setup must install one.
+func TestEnsureKernelInstallsWhenMissing(t *testing.T) {
+	rec := &cmdRec{}
+	s := newCleanupSetup(t, rec, true) // Home is a temp dir: no kernel present
+
+	s.ensureKernel()
+
+	if !rec.ran("container system kernel set --recommended") {
+		t.Errorf("expected the recommended kernel to be installed; calls=%v", rec.calls)
+	}
+}
+
+// When a default kernel is already configured, setup must not re-download it.
+func TestEnsureKernelSkipsWhenPresent(t *testing.T) {
+	rec := &cmdRec{}
+	s := newCleanupSetup(t, rec, true)
+
+	kernelPath := s.defaultKernelPath()
+	if err := os.MkdirAll(filepath.Dir(kernelPath), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(kernelPath, []byte("kernel"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	s.ensureKernel()
+
+	if rec.ran("kernel set") {
+		t.Errorf("a configured kernel must not be reinstalled; calls=%v", rec.calls)
+	}
+}
+
 // With OpenShell absent, cleanup skips the gateway/brew steps but still
 // removes the driver service.
 func TestCleanupWithoutOpenShell(t *testing.T) {

internal/hostsetup/setup.go

@@ -9,6 +9,7 @@ import (
 	"os"
 	"os/exec"
 	"path/filepath"
+	"runtime"
 	"strconv"
 	"strings"
 	"time"
@@ -113,6 +114,12 @@ func (s *Setup) Run(ctx context.Context, opts Options) error {
 		}
 	}
 
+	// 1b. Default guest kernel. apple/container cannot boot any VM without
+	// one, and a fresh install (or one whose user data was deleted, e.g. by
+	// `cleanup --all -d`) ships without it — every sandbox create would fail
+	// with "default kernel not configured for architecture".
+	s.ensureKernel()
+
 	// 2. vmnet network + guest-reachable gateway address.
 	gatewayIP, err := s.ensureNetwork(ctx, opts.Network)
 	if err != nil {
@@ -326,6 +333,36 @@ func (s *Setup) removePrerequisites(opts CleanupOptions) {
 	}
 }
 
+// defaultKernelPath is where apple/container records the default guest kernel
+// for this architecture (a symlink into its kernels directory).
+func (s *Setup) defaultKernelPath() string {
+	return filepath.Join(s.Home, "Library", "Application Support", "com.apple.container",
+		"kernels", "default.kernel-"+runtime.GOARCH)
+}
+
+// ensureKernel installs apple/container's recommended guest kernel when no
+// default is configured. Without it every sandbox create fails at image
+// unpack with "default kernel not configured for architecture".
+//
+// This is deliberately non-fatal: setup is the documented repair command, so a
+// transient download failure must not stop it from fixing the rest of the
+// wiring. It warns with the exact manual fallback instead.
+func (s *Setup) ensureKernel() {
+	if _, err := os.Lstat(s.defaultKernelPath()); err == nil {
+		return
+	}
+	s.Log.Info("setup: no default guest kernel configured; installing the recommended one (one-time, ~600 MB)")
+	err := s.ExecStream("container", "system", "kernel", "set", "--recommended")
+	if err == nil {
+		s.Log.Info("setup: default guest kernel installed")
+		return
+	}
+	s.Log.Warn("setup: could not install the recommended guest kernel — SANDBOXES WILL NOT BOOT until one is set",
+		"err", err,
+		"retry", "container system kernel set --recommended",
+		"fallback", "download the kata-static arm64 tarball with curl, then: container system kernel set --arch arm64 --tar <file> --binary opt/kata/share/kata-containers/vmlinux-<version>")
+}
+
 func (s *Setup) ensureNetwork(ctx context.Context, name string) (string, error) {
 	find := func() (string, error) {
 		networks, err := s.RT.Networks(ctx)