diff --git a/internal/apply/maintenance_window_test.go b/internal/apply/maintenance_window_test.go
new file mode 100644
index 0000000..1649207
--- /dev/null
+++ b/internal/apply/maintenance_window_test.go
@@ -0,0 +1,196 @@
+package apply
+
+import (
+	"context"
+	"errors"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-host/internal/declaration"
+)
+
+// A scheduled step may hold its module's own containers still while it runs (novox/hq ADR 0189,
+// issue 108).
+//
+// What it exists for: the artifact store's collector walks the storage and requires every writer
+// stopped. A run-once step runs beside containers and a scheduled one is the same container again,
+// so the mesh had no way to say it — which is why the store it inherited has never collected
+// anything. The risk the field brings is one shape only: a window that opens and never closes.
+// Every test here is about that shape.
+
+// windowRun records the order of stop / run / start, which is the whole of what is being asserted.
+type windowRun struct {
+	mu     sync.Mutex
+	order  []string
+	failAt string // the arg[0] that should fail ("run" makes the step fail)
+	wontGo string // a container name that refuses to start again
+}
+
+func (w *windowRun) run(_ context.Context, _ string, args ...string) (string, error) {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	switch args[0] {
+	case "info":
+		return "27.0\n", nil
+	case "stop", "start":
+		w.order = append(w.order, args[0]+" "+args[1])
+		if args[0] == "start" && args[1] == w.wontGo {
+			return "", errors.New("the runtime refused")
+		}
+	case "run":
+		w.order = append(w.order, "run")
+		if w.failAt == "run" {
+			return "", errors.New("the step exited non-zero")
+		}
+	}
+	return "", nil
+}
+
+func (w *windowRun) seen() []string {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	return append([]string{}, w.order...)
+}
+
+// aStoreWithACollector is a module in the shape distribution has: a server that must not be
+// writing, and a nightly step that walks its storage with the server held still.
+func aStoreWithACollector(t *testing.T) *declaration.Declaration {
+	t.Helper()
+	return parseTrusted(t, `{"declaration":1,"resources":[
+	  {"id":"store","type":"container","name":"mesh-registry","image":"`+pinned+`"},
+	  {"id":"collect","type":"container","name":"mesh-registry-collect","image":"`+pinned+`",
+	   "schedule":"30 3 * * *","while-stopped":["store"]}
+	]}`)
+}
+
+func fireOnce(t *testing.T, d *declaration.Declaration, w *windowRun) {
+	t.Helper()
+	clock := &fixedClock{now: time.Date(2026, 10, 2, 3, 29, 0, 0, time.UTC)}
+	s := NewScheduler(clock, w.run, func(string) {})
+	s.Sync(d, nil)
+	s.Advance(context.Background(), time.Date(2026, 10, 2, 3, 30, 5, 0, time.UTC))
+	s.Wait()
+}
+
+func TestAScheduledStepHoldsItsModulesContainerStillAndStartsItAgain(t *testing.T) {
+	w := &windowRun{}
+	fireOnce(t, aStoreWithACollector(t), w)
+
+	got := w.seen()
+	want := []string{"stop mesh-registry", "run", "start mesh-registry"}
+	var kept []string
+	for _, line := range got {
+		if strings.HasPrefix(line, "stop mesh-registry-collect") {
+			// Clearing the step's own exited container by name; not part of the window.
+			continue
+		}
+		kept = append(kept, line)
+	}
+	if len(kept) != len(want) {
+		t.Fatalf("the window was not stop, run, start: %v", got)
+	}
+	for i := range want {
+		if kept[i] != want[i] {
+			t.Fatalf("the window was %v, want %v", kept, want)
+		}
+	}
+}
+
+// The one that matters: a step that fails must leave the service running.
+func TestAFailedStepStillClosesTheWindow(t *testing.T) {
+	w := &windowRun{failAt: "run"}
+	fireOnce(t, aStoreWithACollector(t), w)
+
+	var started bool
+	for _, line := range w.seen() {
+		if line == "start mesh-registry" {
+			started = true
+		}
+	}
+	if !started {
+		t.Fatalf("the step failed and the container it held still was never started again: %v", w.seen())
+	}
+}
+
+// A container that will not come back is said loudly: it is down, and nothing else notices until
+// the next apply compares it.
+func TestAContainerThatWillNotStartAgainIsSaidLoudly(t *testing.T) {
+	w := &windowRun{wontGo: "mesh-registry"}
+	var said []string
+	clock := &fixedClock{now: time.Date(2026, 10, 2, 3, 29, 0, 0, time.UTC)}
+	s := NewScheduler(clock, w.run, func(line string) { said = append(said, line) })
+	s.Sync(aStoreWithACollector(t), nil)
+	s.Advance(context.Background(), time.Date(2026, 10, 2, 3, 30, 5, 0, time.UTC))
+	s.Wait()
+
+	var loud bool
+	for _, line := range said {
+		if strings.Contains(line, "WILL NOT START AGAIN") && strings.Contains(line, "mesh-registry") {
+			loud = true
+		}
+	}
+	if !loud {
+		t.Fatalf("a service left stopped by a maintenance window was not said loudly: %v", said)
+	}
+}
+
+// Several containers come back in the reverse of the order they were stopped: a module names the
+// dependant first, and starting it before what it depends on is not bringing it back.
+func TestTheWindowClosesInTheReverseOfTheOrderItOpened(t *testing.T) {
+	d := parseTrusted(t, `{"declaration":1,"resources":[
+	  {"id":"web","type":"container","name":"web","image":"`+pinned+`"},
+	  {"id":"db","type":"container","name":"db","image":"`+pinned+`"},
+	  {"id":"collect","type":"container","name":"collect","image":"`+pinned+`",
+	   "schedule":"30 3 * * *","while-stopped":["web","db"]}
+	]}`)
+	w := &windowRun{}
+	fireOnce(t, d, w)
+
+	var stops, starts []string
+	for _, line := range w.seen() {
+		switch {
+		case line == "stop web" || line == "stop db":
+			stops = append(stops, line)
+		case strings.HasPrefix(line, "start "):
+			starts = append(starts, line)
+		}
+	}
+	if len(stops) != 2 || stops[0] != "stop web" || stops[1] != "stop db" {
+		t.Fatalf("stopped in %v, want the order the step named them", stops)
+	}
+	if len(starts) != 2 || starts[0] != "start db" || starts[1] != "start web" {
+		t.Fatalf("started in %v, want the reverse", starts)
+	}
+}
+
+// And the refusals, each for what it says rather than that it says something.
+func TestAMaintenanceWindowIsRefusedWhereItCannotMean(t *testing.T) {
+	for _, c := range []struct{ name, body, says string }{
+		{
+			"a window with no schedule",
+			`{"id":"collect","type":"container","name":"c","image":"` + pinned + `","while-stopped":["store"]}`,
+			"needs a schedule",
+		},
+		{
+			"a window naming itself",
+			`{"id":"collect","type":"container","name":"c","image":"` + pinned + `","schedule":"30 3 * * *","while-stopped":["collect"]}`,
+			"this step itself",
+		},
+		{
+			"a window naming something that is not a container here",
+			`{"id":"collect","type":"container","name":"c","image":"` + pinned + `","schedule":"30 3 * * *","while-stopped":["elsewhere"]}`,
+			"no container by that id",
+		},
+	} {
+		_, err := declaration.ParseTrusted([]byte(`{"declaration":1,"resources":[` + c.body + `]}`))
+		if err == nil {
+			t.Errorf("%s was accepted", c.name)
+			continue
+		}
+		if !strings.Contains(err.Error(), c.says) {
+			t.Errorf("%s: the refusal does not say %q: %v", c.name, c.says, err)
+		}
+	}
+}
diff --git a/internal/apply/schedule.go b/internal/apply/schedule.go
index 6c80de1..179d3c4 100644
--- a/internal/apply/schedule.go
+++ b/internal/apply/schedule.go
@@ -65,6 +65,29 @@ type scheduledJob struct {
 	container *declaration.Container
 	next      time.Time // the next minute at which it is due
 	running   bool      // a run is in flight — the next due run is skipped rather than stacked
+	// hold is the runtime names of the containers held still for the duration of a run, in the
+	// order the step named them (novox/hq ADR 0189).
+	hold []string
+}
+
+// heldStillFor is the runtime names of the containers a step holds still, resolved from ids.
+func heldStillFor(step *declaration.Container, d *declaration.Declaration) []string {
+	if len(step.WhileStopped) == 0 {
+		return nil
+	}
+	byID := map[string]string{}
+	for _, r := range d.Resources {
+		if c, ok := r.(*declaration.Container); ok {
+			byID[c.Identity()] = c.Name
+		}
+	}
+	out := make([]string, 0, len(step.WhileStopped))
+	for _, id := range step.WhileStopped {
+		if name := byID[id]; name != "" {
+			out = append(out, name)
+		}
+	}
+	return out
 }
 
 // NewScheduler builds a scheduler. A nil clock is the system clock; a nil log says nothing.
@@ -123,15 +146,20 @@ func (s *Scheduler) Sync(d *declaration.Declaration, held map[string]bool) {
 		// on its cadence rather than staying running to be restarted — so its identity cannot
 		// depend on another resource's content and there is nothing to pass.
 		spec := containerSpec(c, inputs{})
+		// The runtime stops containers by name; the declaration names them by id. Resolved here,
+		// against the declaration this job was armed from, so a fire never has to look anything up
+		// (novox/hq ADR 0189). The parser has already refused an id that is not a container here.
+		hold := heldStillFor(c, d)
 		if existing := s.jobs[c.Identity()]; existing != nil && existing.spec == spec {
 			// Unchanged: keep where it is in its cadence, refresh the declaration pointer only.
 			existing.container = c
+			existing.hold = hold
 			continue
 		}
 		// New or changed: arm it for the next due minute after now.
 		next, _ := cron.Next(s.clock.Now())
 		s.jobs[c.Identity()] = &scheduledJob{
-			id: c.Identity(), spec: spec, cron: cron, container: c, next: next,
+			id: c.Identity(), spec: spec, cron: cron, container: c, next: next, hold: hold,
 		}
 	}
 
@@ -202,6 +230,15 @@ func (s *Scheduler) fire(ctx context.Context, j *scheduledJob) {
 		return
 	}
 
+	// **The window opens here and closes in the defer, whatever happens** (novox/hq ADR 0189).
+	// Deferred before the first stop so a panic, a failing step or a step that runs long all end
+	// the same way: the service running. The one real risk of this field is a window that never
+	// closes, and the only defence against it is that closing is not conditional on anything.
+	if len(j.hold) > 0 {
+		defer s.letRun(ctx, cri, j)
+		s.holdStill(ctx, cri, j)
+	}
+
 	// A container by this name left exited by the previous run would collide with --name. Removing
 	// one that is not there is the state we want, so its error is ignored — the same as run-once.
 	_, _ = s.run(ctx, cri, "rm", "-f", j.container.Name)
@@ -260,3 +297,51 @@ func (s *Scheduler) Run(ctx context.Context) {
 		}
 	}
 }
+
+// holdStill stops the containers this step runs instead of, in the order it named them.
+//
+// A stop that fails is said and not fatal. The step runs anyway: for the case this exists for —
+// a collector walking storage nothing must be writing to — a writer that would not stop is worth
+// knowing about, and refusing to run would mean the work never happens and the log says nothing
+// new each night. What must not be skipped is the restart, and it is not: it is deferred.
+func (s *Scheduler) holdStill(ctx context.Context, cri string, j *scheduledJob) {
+	for _, name := range j.hold {
+		if _, err := s.run(ctx, cri, "stop", name); err != nil {
+			s.log(fmt.Sprintf("scheduled step %s: could not stop %s for the run: %v", j.id, name, err))
+			continue
+		}
+		s.log(fmt.Sprintf("scheduled step %s: %s held still for the run", j.id, name))
+	}
+}
+
+// letRun starts them again, in the reverse of the order they were stopped, and says so loudly if
+// one does not come back.
+//
+// **Reverse order**, because stopping walks a dependency the other way: a module that holds two
+// containers still names the one that depends on the other first, and bringing them back the same
+// way would start a dependant before what it depends on.
+//
+// Given its own context, because this runs in a defer and the one the run used may already be
+// cancelled — a host shutting down mid-window would otherwise leave the service stopped, which is
+// precisely the outcome this field must never have.
+func (s *Scheduler) letRun(_ context.Context, cri string, j *scheduledJob) {
+	ctx, cancel := context.WithTimeout(context.Background(), closingWindow)
+	defer cancel()
+	for i := len(j.hold) - 1; i >= 0; i-- {
+		name := j.hold[i]
+		if _, err := s.run(ctx, cri, "start", name); err != nil {
+			// Said as loudly as this host says anything: a service the mesh stopped for a
+			// maintenance window and could not start again is down, and nothing else will notice
+			// until the next apply compares it.
+			s.log(fmt.Sprintf(
+				"scheduled step %s: %s was held still for the run and WILL NOT START AGAIN: %v",
+				j.id, name, err))
+			continue
+		}
+		s.log(fmt.Sprintf("scheduled step %s: %s running again", j.id, name))
+	}
+}
+
+// closingWindow is how long the host will spend putting back what it stopped. Generous: this is
+// the half that must not be given up on.
+const closingWindow = 5 * time.Minute
diff --git a/internal/declaration/declaration.go b/internal/declaration/declaration.go
index c7419b4..020e105 100644
--- a/internal/declaration/declaration.go
+++ b/internal/declaration/declaration.go
@@ -997,6 +997,24 @@ type Container struct {
 	// rather than stacked. It is exclusive with RunOnce and with restart-on: a container runs once
 	// and gates, runs on a cadence, or stays up — never two of these.
 	Schedule string `json:"schedule,omitempty"`
+
+	// WhileStopped names resources of the same module — containers — that must be held still for
+	// the duration of this step's run (novox/hq ADR 0189). The host stops each before the run and
+	// starts each again after it, **whatever the step did**: a step that failed must leave the
+	// service running, because the one real risk of this field is a window that never closes.
+	//
+	// **For the work a service cannot have done underneath it.** The artifact store's collector
+	// walks the storage and requires every writer stopped; a run-once step runs beside containers
+	// and a scheduled one is the same container again, so until this there was no way for a module
+	// to say it. The predecessor said it with a shell script, which is how the mesh inherited a
+	// store that has never collected anything.
+	//
+	// **Its own module's containers, and only on a schedule.** A module that could quiesce a
+	// neighbour could stop the mesh. And at apply time the host already has a window — the
+	// declaration is applied in order and a run-once step gates what follows — so a one-time
+	// offline job says *before*, not *instead of*; a recurring window is the case order cannot
+	// express, and the only one this serves.
+	WhileStopped []string `json:"while-stopped,omitempty"`
 }
 
 func (c *Container) Identity() string { return c.ID }
@@ -1028,6 +1046,20 @@ func (c *Container) validate(where string, _ bool) []string {
 			problems = append(problems, where+": "+err.Error())
 		}
 	}
+	// A maintenance window belongs to a recurring step (novox/hq ADR 0189). Refused on anything
+	// else here, where the field is; that it names containers of the same module, and not itself,
+	// is judged against the whole declaration (see whileStoppedNames).
+	if len(c.WhileStopped) > 0 && c.Schedule == "" {
+		problems = append(problems, where+": while-stopped needs a schedule; at apply the host "+
+			"already has a window — the declaration is applied in order and a run-once step gates "+
+			"what follows — so a one-time offline job is declared before what it works on")
+	}
+	for _, id := range c.WhileStopped {
+		if id == c.ID {
+			problems = append(problems, where+": while-stopped names "+strconv.Quote(id)+
+				", which is this step itself")
+		}
+	}
 	// The runtime's flags take addresses, and a name here would be handed to it verbatim and
 	// refused at create — after the old container was already removed. Refused on arrival instead.
 	for _, d := range c.Dns {
@@ -1438,6 +1470,7 @@ func parse(raw []byte, allowActions bool) (*Declaration, error) {
 		problems = append(problems, resource.validate(where, allowActions)...)
 		d.Resources = append(d.Resources, resource)
 	}
+	problems = append(problems, checkWhileStopped(d.Resources)...)
 	problems = append(problems, checkAdoption(env.Adoption, d.Resources, allowActions)...)
 	if env.Adoption == nil {
 		for _, r := range d.Resources {
@@ -1466,6 +1499,41 @@ func parse(raw []byte, allowActions bool) (*Declaration, error) {
 	return d, nil
 }
 
+// checkWhileStopped judges a maintenance window against the whole declaration (novox/hq ADR 0189).
+//
+// A step may hold still only a container that is **here** — in this same declaration, which is to
+// say on this machine and placed by the mesh. That is what makes it the module's own: a node's
+// declaration carries one module's resources beside another's, so the id must also be a container
+// and not a file or a directory, which there would be nothing to stop.
+//
+// Refused on arrival rather than discovered at the first fire. A window that names something the
+// host cannot stop is a window that opens at 03:00 and reports nothing until somebody reads a log.
+func checkWhileStopped(resources []Resource) []string {
+	containers := map[string]bool{}
+	for _, r := range resources {
+		if r.Kind() == TypeContainer {
+			containers[r.Identity()] = true
+		}
+	}
+	var problems []string
+	for _, r := range resources {
+		c, ok := r.(*Container)
+		if !ok {
+			continue
+		}
+		for _, id := range c.WhileStopped {
+			if containers[id] {
+				continue
+			}
+			problems = append(problems, fmt.Sprintf(
+				"resource %q: while-stopped names %q, and this declaration has no container by "+
+					"that id. A step may hold still only a container placed on this machine "+
+					"beside it", c.ID, id))
+		}
+	}
+	return problems
+}
+
 func strictDecode(raw []byte, into any) error {
 	// DisallowUnknownFields is the whole point rather than strictness for its own sake: a
 	// field the host does not know is a thing the control plane believes it asked for.