Two machines holding the build role share one queue, and neither is handed an ask while busy (hq ADR 0190)

Against a real bus: three asks, two machines; each takes one, the third waits until one is free
and then goes to that one; a machine that stops leaves nothing taken twice. The redelivery of an
ask a dead machine held is the ack wait's, proven by the hand-back test beside this one.

And the order a live mesh switches over in, written where the role is named: queued builds first,
then this controller, then build-agent assigned where machines build, then the builder and the old
seat's stream forgotten.

internal/broker/derived.go

@@ -144,12 +144,18 @@ func ConsumerFor(p Principal) (Consumer, bool) {
 	}, true
 }
 
-// HolderConsumerFor is the worker a seat's holder gets on that seat's work queue.
+// HolderConsumerFor is the worker a seat's holders share on that seat's work queue.
 //
-// **A queue group even though the seat guarantees one holder.** The seat is *authority* — who may
-// be the telegram sender — and the queue group is *delivery*. Tie delivery to the seat and the
-// day somebody allows two holders for throughput, every message is processed twice with nothing
-// reporting it. Kept separate, relaxing one changes nothing about the other.
+// **One worker for every holder, and each holder pulls one ask when it is idle** (novox/hq ADR
+// 0190). The seat is *authority* — who may be the telegram sender — and the worker is *delivery*,
+// kept separate so that relaxing one changes nothing about the other: a node-scoped seat has a
+// holder per machine, and all of them take from this one consumer, so the work is shared without
+// any holder knowing about the others. Pulled rather than pushed because a push consumer hands the
+// next ask to whichever subscriber the server picks, busy or not, and a pulled one is asked for by
+// a holder that has just become free. Which is also what ends the race issue 186 describes — asks
+// delivered behind the one being worked, expiring unacknowledged and dropped after the fifth
+// redelivery: nothing is delivered that nobody asked for. A long build keeps its own ask alive
+// (stillWorking); the ack wait is for a holder that died.
 func HolderConsumerFor(node, module string, seat DeclaredSeat) (Consumer, bool) {
 	if len(seat.Accepts) == 0 {
 		return Consumer{}, false
@@ -158,18 +164,13 @@ func HolderConsumerFor(node, module string, seat DeclaredSeat) (Consumer, bool)
 		Name:           "SEAT_" + upperSnake(seat.Name) + "_worker",
 		Stream:         seatStreamName(seat.Name),
 		Filters:        []string{"mesh.seat." + seat.Name + ".accept.>"},
-		Queue:          "holders",
 		AckWaitSeconds: 60,
 		MaxDeliver:     5,
-		// **One in flight.** A holder works one ask at a time, so the server hands it one at a
-		// time: with the default of many, every ask behind the one being worked was delivered,
-		// left unacknowledged for the length of the work, redelivered after the ack wait, and
-		// after the fifth time dropped — on 2026-10-01 twenty-six of forty-three builds asked in
-		// two minutes were never built, and the queue read as empty (novox/hq issue 186).
-		MaxAckPending: 1,
-		Why: fmt.Sprintf("%s on %s holds %s; it acknowledges after the work is done, so a "+
-			"crash mid-work redelivers rather than loses; one in flight, so a queue of asks is a "+
-			"queue and not a race against the ack wait", module, node, seat.Name),
+		// As many in flight as there are holders working, which pulling bounds by itself: a holder
+		// fetches one and fetches again only after it acknowledged. The server's default stands.
+		Why: fmt.Sprintf("%s on %s holds %s; every holder pulls one ask at a time from this worker "+
+			"and acknowledges after the work is done, so a crash mid-work redelivers rather than "+
+			"loses and an idle holder is the one that takes the next ask", module, node, seat.Name),
 	}, true
 }
 

internal/broker/derived_test.go

@@ -88,15 +88,20 @@ func TestAModuleThatConsumesNothingGetsNoConsumer(t *testing.T) {
 	}
 }
 
-// The seat is authority and the queue group is delivery. Tie them together and the day somebody
-// allows two holders, every message is processed twice with nothing reporting it.
-func TestAHoldersWorkerUsesAQueueGroupAnyway(t *testing.T) {
+// The seat is authority and the worker is delivery (novox/hq ADR 0190): one worker per seat, shared
+// by every holder and pulled from, so a second holder takes the next ask rather than a copy of the
+// same one — which is what a queue group used to guard, and what pulling one durable gives outright.
+func TestAHoldersWorkerIsOneSharedByItsHolders(t *testing.T) {
 	c, ok := HolderConsumerFor("one", "telegram", telegramSeat())
 	if !ok {
 		t.Fatal("the holder of a seat with inbound work got no worker")
 	}
-	if c.Queue == "" {
-		t.Fatal("the worker is not in a queue group, so a second holder would double-process")
+	two, _ := HolderConsumerFor("two", "telegram", telegramSeat())
+	if c.Name != two.Name || c.Stream != two.Stream {
+		t.Fatal("two holders got two workers, so each would process every ask")
+	}
+	if c.Push || c.Queue != "" {
+		t.Fatal("the worker is pushed, so the server would hand an ask to a busy holder")
 	}
 	if c.Stream != "SEAT_TELEGRAM_SENDER" {
 		t.Fatalf("the worker reads %q, not the seat's own stream", c.Stream)
@@ -154,15 +159,23 @@ func TestANodesDeclarationConsumerIsWhatItsOwnGrantAllows(t *testing.T) {
 	has(t, perms.Subscribe, c.Filters[0])
 }
 
-// A holder works one ask at a time, so the server hands it one at a time (novox/hq issue 186):
-// asks queued behind the one being worked wait in the stream rather than being delivered,
-// left to expire and dropped after the fifth redelivery.
-func TestAHoldersWorkerTakesOneAskAtATime(t *testing.T) {
-	c, found := HolderConsumerFor("anchor", "builder", DeclaredSeat{Name: "mesh-build-machine", Accepts: []string{"build"}})
+// Every holder of a seat shares one worker and pulls from it (novox/hq ADR 0190): no queue group
+// and no delivery subject, because a push consumer hands the next ask to whichever subscriber the
+// server picks, busy or not; and no cap of one in flight, because pulling bounds the asks in flight
+// by the holders that are free — which is what ended the race of issue 186, where asks delivered
+// behind the one being worked expired and were dropped.
+func TestAHoldersWorkerIsPulledByEveryHolder(t *testing.T) {
+	c, found := HolderConsumerFor("anchor", "build-agent", DeclaredSeat{Name: "node-build-agent", Accepts: []string{"build"}})
 	if !found {
 		t.Fatal("a seat that accepts work has no worker")
 	}
-	if c.MaxAckPending != 1 {
-		t.Fatalf("the worker may have %d asks in flight; one, so a queue is a queue", c.MaxAckPending)
+	if c.Queue != "" || c.Push {
+		t.Fatalf("the worker is pushed (queue %q, push %v); a holder pulls when it is free", c.Queue, c.Push)
+	}
+	if c.MaxAckPending != 0 {
+		t.Fatalf("the worker caps asks in flight at %d; pulling bounds them by the holders working", c.MaxAckPending)
+	}
+	if c.Name != "SEAT_NODE_BUILD_AGENT_worker" || c.Stream != "SEAT_NODE_BUILD_AGENT" {
+		t.Fatalf("the worker is %s on %s; one per seat, shared by its holders", c.Name, c.Stream)
 	}
 }

internal/broker/nats.go

@@ -110,10 +110,10 @@ type Principal struct {
 	PasswordHash string
 }
 
-// meshSeatsTheControllerUses are the roles the mesh's own flows submit work to. Named rather than
+// seatsTheControllerAsks are the roles the mesh's own flows submit work to. Named rather than
 // derived from the seat set: the controller is not a module and declares no `uses`, so its side of a
 // seat has to be stated, and a list is what makes "which roles does the mesh itself talk to" answerable.
-var meshSeatsTheControllerUses = []string{"mesh-build-machine"}
+var seatsTheControllerAsks = []string{"node-build-agent"}
 
 // enrolmentPrefix is the space every enrolling node's user and inbox live under, so the one place the
 // controller may answer an enrolment is derived from the same constant the user is named from.
@@ -208,7 +208,9 @@ func PermissionsFor(p Principal) (Permissions, error) {
 		// Work the mesh's own flows submit to a role, and the outcomes they wait on (ADR 0121). A
 		// build is the one today: the controller asks, and reads the answer from the seat's event
 		// like the catalogue does — which is why no holder needs to publish into anybody's inbox.
-		for _, seat := range meshSeatsTheControllerUses {
+		// A node-scoped seat's work subject carries no node (novox/hq ADR 0190): the ask goes to
+		// the role, and whichever machine holding it is idle takes it.
+		for _, seat := range seatsTheControllerAsks {
 			pub = append(pub, "mesh.seat."+seat+".accept.>")
 		}
 		// **And what the mesh says it did** (novox/hq ADR 0134). The control plane states its own
@@ -368,13 +370,17 @@ func PermissionsFor(p Principal) (Permissions, error) {
 
 		// 3. Seats it holds: full participation.
 		for _, s := range p.Holds {
-			// Taking work from the role's queue: the worker consumer it binds (asked about,
-			// delivered on, acknowledged), each on the seat's own stream. The first machine to
-			// take work over the new bus was refused the asking (2026-09-28).
+			// Taking work from the role's queue: the worker consumer every holder shares (asked
+			// about, pulled from, acknowledged), on the seat's own stream (novox/hq ADR 0190). A
+			// holder pulls — asks the consumer for its next message, answered on its own inbox —
+			// so what it needs is MSG.NEXT on that worker and nothing delivered to it. The first
+			// machine to take work over the new bus was refused the asking (2026-09-28).
 			worker := "SEAT_" + upperSnake(s.Name) + "_worker"
 			stream := seatStreamName(s.Name)
-			sub = append(sub, "_DELIVER."+worker, "_DELIVER."+worker+".>")
-			pub = append(pub, "$JS.API.CONSUMER.INFO."+stream+"."+worker, "$JS.ACK."+stream+"."+worker+".>")
+			pub = append(pub,
+				"$JS.API.CONSUMER.INFO."+stream+"."+worker,
+				"$JS.API.CONSUMER.MSG.NEXT."+stream+"."+worker,
+				"$JS.ACK."+stream+"."+worker+".>")
 			for _, a := range s.Accepts {
 				sub = append(sub, seatSubject(s, "accept", a))
 			}

internal/broker/nats_test.go

@@ -441,3 +441,24 @@ func contains(list []string, want string) bool {
 	}
 	return false
 }
+
+// A node-scoped seat's work is shared (novox/hq ADR 0190): its holder on any machine subscribes the
+// seat's one work subject, with no node in it, so holders on several machines read one queue. The
+// node token belongs to a seat's tools, which are asked of one machine (design 33 §4), not to its work.
+func TestANodeSeatsWorkSubjectCarriesNoNode(t *testing.T) {
+	seat := Seat{Name: "node-build-agent", Scope: "node", Accepts: []string{"build"}, Serves: []string{"status"}}
+	perms, err := PermissionsFor(Principal{Kind: KindModule, Node: "anchor", Module: "build-agent", Holds: []Seat{seat}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	has(t, perms.Subscribe, "mesh.seat.node-build-agent.accept.build")
+	hasNot(t, perms.Subscribe, "mesh.seat.node-build-agent.accept.build.anchor")
+	// And its tools still carry the machine.
+	has(t, perms.Subscribe, "mesh.seat.node-build-agent.tool.status.anchor")
+	// The controller asks the role, not a machine.
+	controller, err := PermissionsFor(Principal{Kind: KindController})
+	if err != nil {
+		t.Fatal(err)
+	}
+	has(t, controller.Publish, "mesh.seat.node-build-agent.accept.>")
+}

internal/broker/streams.go

@@ -217,7 +217,7 @@ var ControllerFollows = []string{
 	// A build's outcome, which is the build-machine role's own event now (ADR 0121) rather than a
 	// message on the control branch. Same three audiences, one publish: whoever asked, this, and the
 	// catalogue.
-	seatEventSubject("mesh-build-machine", "built"),
+	seatEventSubject("node-build-agent", "built"),
 	// The forge's merges: what moved a source, so the mesh builds what that source produces
 	// without anybody telling it (novox/hq 04-ISSUES/131). Appended, because the index is a name.
 	moduleEventSubject("gitea", "pull.merged"),

internal/broker/testdata/composed.conf

@@ -24,8 +24,8 @@ accounts {
     jetstream: enabled
     users = [
       { user: "controller", password: "$2a$11$cccccccccccccccccccccc", permissions: {
-        publish: { allow: ["$JS.ACK.CONTROL.controller.>", "$JS.ACK.EVENTS.controller.>", "$JS.API.>", "_INBOX.enrol.>", "mesh.assignment.>", "mesh.control.>", "mesh.mod.*.tool.>", "mesh.node.>", "mesh.seat.mesh-build-machine.accept.>", "mesh.seat.mesh-controller.event.applied", "mesh.seat.mesh-controller.event.built-before", "mesh.seat.mesh-controller.event.refused"] }
-        subscribe: { allow: ["$JS.API.>", "_DELIVER.controller", "_DELIVER.controller.>", "_INBOX.controller.>", "mesh.control.>", "mesh.mod.gitea.event.pull.merged", "mesh.mod.mesh-catalog.event.catching-up", "mesh.mod.mesh-catalog.event.upgraded", "mesh.seat.mesh-build-machine.event.built", "mesh.seat.mesh-controller.tool.>"] }
+        publish: { allow: ["$JS.ACK.CONTROL.controller.>", "$JS.ACK.EVENTS.controller.>", "$JS.API.>", "_INBOX.enrol.>", "mesh.assignment.>", "mesh.control.>", "mesh.mod.*.tool.>", "mesh.node.>", "mesh.seat.mesh-controller.event.applied", "mesh.seat.mesh-controller.event.built-before", "mesh.seat.mesh-controller.event.refused", "mesh.seat.node-build-agent.accept.>"] }
+        subscribe: { allow: ["$JS.API.>", "_DELIVER.controller", "_DELIVER.controller.>", "_INBOX.controller.>", "mesh.control.>", "mesh.mod.gitea.event.pull.merged", "mesh.mod.mesh-catalog.event.catching-up", "mesh.mod.mesh-catalog.event.upgraded", "mesh.seat.mesh-controller.tool.>", "mesh.seat.node-build-agent.event.built"] }
         allow_responses: { max: 1, ttl: "1m" }
       } }
       { user: "enrol.one", password: "$2a$11$eeeeeeeeeeeeeeeeeeeeee", permissions: {
@@ -37,8 +37,8 @@ accounts {
         subscribe: { allow: ["_DELIVER.one", "_DELIVER.one.>", "_INBOX.node.one.>", "mesh.node.one.declare"] }
       } }
       { user: "one.telegram", password: "$2a$11$tttttttttttttttttttttt", permissions: {
-        publish: { allow: ["$JS.ACK.EVENTS.one_telegram.>", "$JS.ACK.SEAT_TELEGRAM_SENDER.SEAT_TELEGRAM_SENDER_worker.>", "$JS.API.CONSUMER.INFO.EVENTS.one_telegram", "$JS.API.CONSUMER.INFO.SEAT_TELEGRAM_SENDER.SEAT_TELEGRAM_SENDER_worker", "$JS.API.CONSUMER.MSG.NEXT.EVENTS.one_telegram", "$JS.API.DIRECT.GET.ASSIGNMENTS.mesh.assignment.one.telegram", "mesh.seat.telegram-sender.event.delivered", "mesh.seat.telegram-sender.event.failed"] }
-        subscribe: { allow: ["_DELIVER.SEAT_TELEGRAM_SENDER_worker", "_DELIVER.SEAT_TELEGRAM_SENDER_worker.>", "_INBOX.one.telegram.>", "mesh.assignment.one.telegram", "mesh.mod.telegram.tool.>", "mesh.seat.telegram-sender.accept.send"] }
+        publish: { allow: ["$JS.ACK.EVENTS.one_telegram.>", "$JS.ACK.SEAT_TELEGRAM_SENDER.SEAT_TELEGRAM_SENDER_worker.>", "$JS.API.CONSUMER.INFO.EVENTS.one_telegram", "$JS.API.CONSUMER.INFO.SEAT_TELEGRAM_SENDER.SEAT_TELEGRAM_SENDER_worker", "$JS.API.CONSUMER.MSG.NEXT.EVENTS.one_telegram", "$JS.API.CONSUMER.MSG.NEXT.SEAT_TELEGRAM_SENDER.SEAT_TELEGRAM_SENDER_worker", "$JS.API.DIRECT.GET.ASSIGNMENTS.mesh.assignment.one.telegram", "mesh.seat.telegram-sender.event.delivered", "mesh.seat.telegram-sender.event.failed"] }
+        subscribe: { allow: ["_INBOX.one.telegram.>", "mesh.assignment.one.telegram", "mesh.mod.telegram.tool.>", "mesh.seat.telegram-sender.accept.send"] }
         allow_responses: { max: 1, ttl: "1m" }
       } }
       { user: "two.audit", password: "$2a$11$aaaaaaaaaaaaaaaaaaaaaa", permissions: {

internal/catalogue/seats.go

@@ -96,8 +96,17 @@ var defaultSeats = []Seat{
 	// A build says what it does as it does it (novox/hq ADR 0157): `started` when work is taken,
 	// `log.<build id>` for every line, `built` for the outcome. The log's tail token is the build's
 	// id, so a reader follows one build by subject alone.
+	// **Node-scoped, and every holder takes from one queue** (novox/hq ADR 0190): a build is asked of
+	// the role, and whichever machine holding the seat is idle pulls it. One holder per machine is
+	// what the scope says; sharing the work is what a seat's queue has always done.
+	{Name: "node-build-agent", Scope: ScopeNode,
+		Accepts: []string{"build"}, Emits: []string{"started", "built", "log.*"}, Decision: "novox/hq ADR 0190"},
+	// **Retired by ADR 0190, kept while a manifest still claims it.** The one build machine's seat.
+	// A claim to a seat the mesh no longer defines is refused, and the module holding this one is
+	// assigned on a live machine until build-agent replaces it — removing the row first would make
+	// that machine unresolvable in the meantime. Deleted once no registered manifest claims it.
 	{Name: "mesh-build-machine", Scope: ScopeMesh,
-		Accepts: []string{"build"}, Emits: []string{"started", "built", "log.*"}, Decision: "novox/hq ADR 0121"},
+		Accepts: []string{"build"}, Emits: []string{"started", "built", "log.*"}, Decision: "novox/hq ADR 0190"},
 	{Name: "node-dns-resolver", Scope: ScopeNode, Decision: "novox/hq ADR 0121"},
 	// The intrusion prevention's verbs (novox/hq ADR 0179): what a person asks a machine's ban list
 	// whatever keeps it — who is banned and why, ban one address, let one go. Every holder serves all

internal/catalogue/seats_test.go

@@ -44,9 +44,10 @@ func TestTheSeatsAreAClosedSetAndEachNamesItsDecision(t *testing.T) {
 			delivered[s.Delivers] = s.Name
 		}
 	}
-	// Sixteen since node-service-manager (novox/hq ADR 0177).
-	if len(Seats()) != 16 {
-		t.Errorf("the mesh defines %d seats rather than 16; the set is closed, so a change here is "+
+	// Seventeen since node-build-agent (novox/hq ADR 0190) — sixteen once the retired
+	// mesh-build-machine row goes, when no registered manifest claims it any more.
+	if len(Seats()) != 17 {
+		t.Errorf("the mesh defines %d seats rather than 17; the set is closed, so a change here is "+
 			"a decision (novox/hq ADR 0110): %s", len(Seats()), seatNames())
 	}
 }

internal/inventory/dependencies.go

@@ -63,7 +63,7 @@ func dependenciesOf(entries []Entry, against map[string][]string, read map[strin
 		if r := repositoryKey(e.Source.Repository); r != "" {
 			byRepository[r] = append(byRepository[r], name)
 		}
-		if e.Manifest.ClaimsSeat("mesh-build-machine") {
+		if e.Manifest.ClaimsSeat("node-build-agent") || e.Manifest.ClaimsSeat("mesh-build-machine") {
 			builders = append(builders, name)
 		}
 	}

internal/inventory/dependencies_test.go

@@ -12,7 +12,7 @@ func TestDependenciesAreOneRelationWithTheirKinds(t *testing.T) {
 		return Entry{Manifest: catalogue.Manifest{Module: name}, Source: Source{Repository: repository}}
 	}
 	builder := entry("builder", "http://forge/novox/mesh-catalog.git")
-	builder.Manifest.Claims = []catalogue.Claim{{Name: "mesh-build-machine", Scope: catalogue.ScopeMesh}}
+	builder.Manifest.Claims = []catalogue.Claim{{Name: "node-build-agent", Scope: catalogue.ScopeNode}}
 	plugin := entry("shop-plugin", "http://forge/novox/mesh-catalog.git")
 	plugin.Manifest.Build = &catalogue.Build{On: []catalogue.BuildsOn{{Arg: "BASE", Module: "shop"}}}
 	entries := []Entry{

internal/link/builds.go

@@ -19,8 +19,17 @@ import (
 // act on or a declaration a node reconciles toward; a build is a request that takes minutes and has
 // exactly one answer. Too long for request/reply, too particular to be an event.
 
-// TheBuildMachine is the role a build is submitted to.
-const TheBuildMachine = "mesh-build-machine"
+// TheBuildMachine is the role a build is submitted to: node-scoped, held on every machine that
+// builds, and the work shared among them (novox/hq ADR 0190). The name stays for every caller; what
+// it names moved from the mesh's one build machine to whichever build agent is idle.
+//
+// **Switching a live mesh over, in order.** The old seat's stream and worker
+// (SEAT_MESH_BUILD_MACHINE, SEAT_MESH_BUILD_MACHINE_worker) stay on the bus until removed by hand,
+// and the builder module keeps draining them while it is assigned. From the moment a controller
+// with this name runs, new asks go to node-build-agent and wait in its stream until some machine
+// holds the seat. So: let the queued builds finish; roll this controller; register and assign
+// build-agent to the machines that build; unassign builder and forget it and its seat's stream.
+const TheBuildMachine = "node-build-agent"
 
 // BuildWork is where a build request lands, and BuildOutcome is where its result does. Derived from
 // the seat, so both sides name the role and neither names the other.

internal/link/builds_current_test.go

@@ -26,7 +26,7 @@ func TestTheOldBusAnnouncesABuildUnderBothNames(t *testing.T) {
 	if KeyRoleBuilt != "built" {
 		t.Fatalf("the role's event is %q, and a holder emits its verbs bare", KeyRoleBuilt)
 	}
-	if TheBuildMachine != "mesh-build-machine" {
+	if TheBuildMachine != "node-build-agent" {
 		t.Fatalf("the role is %q", TheBuildMachine)
 	}
 	// The two must differ, or one publish would serve both and this doubling would be pointless.

internal/link/builds_nats.go

@@ -126,29 +126,31 @@ func (m *natsMachine) Close() {
 	}
 }
 
-// Take binds to the role's worker and hands each request over, one at a time.
+// Take binds to the role's worker and pulls one request at a time, handing each over.
 //
 // **Bound, never created.** The work queue and the worker on it are the controller's to define
 // (design 25 §3), and a build machine reaches no part of the JetStream API — so a missing one is said
 // as the mesh's to answer rather than quietly created with whatever this client defaults to.
+//
+// **Pulled, one at a time, by whichever holder is free** (novox/hq ADR 0190). Every machine holding
+// the role binds this same worker; a machine asks for the next request only when it has finished
+// the last, so a slow machine never holds an ask an idle one could take, and a machine that took
+// five at once would run five container builds against one runtime and finish all of them slower
+// than the first.
 func (m *natsMachine) Take(ctx context.Context, do func(context.Context, Build)) error {
-	worker, found := broker.HolderConsumerFor(m.on, "builder",
+	worker, found := broker.HolderConsumerFor(m.on, "build-agent",
 		broker.DeclaredSeat{Name: m.seat, Accepts: []string{"build"}})
 	if !found {
 		return fmt.Errorf("%s accepts no work, so there is nothing for this machine to take", m.seat)
 	}
 
-	// One at a time, which the consumer's own ack-pending limit enforces rather than a prefetch
-	// setting: a machine that took five requests at once would run five container builds against one
-	// runtime and finish all of them slower than the first.
-	work := make(chan *nats.Msg, 1)
 	// **The consumer's own filter, not the one subject this machine cares about.** The client checks
 	// what is asked for against the consumer's filter and refuses anything that is not the same —
 	// "subject does not match consumer" — so subscribing `…accept.build` against a consumer filtered
 	// on `…accept.>` is rejected even though it is narrower. Learned twice now, on two different
 	// consumers, which is why it is written down here.
 	filter := worker.Filters[0]
-	sub, err := m.js.Context().ChanQueueSubscribe(filter, worker.Queue, work,
+	sub, err := m.js.Context().PullSubscribe(filter, worker.Name,
 		nats.Bind(worker.Stream, worker.Name), nats.ManualAck())
 	if err != nil {
 		return fmt.Errorf(
@@ -159,13 +161,27 @@ func (m *natsMachine) Take(ctx context.Context, do func(context.Context, Build))
 	m.sub = sub
 
 	for {
-		select {
-		case <-ctx.Done():
+		if ctx.Err() != nil {
 			return nil
-		case msg, ok := <-work:
-			if !ok {
-				return errors.New("the bus stopped delivering build work")
+		}
+		// One, and wait a while for it; an empty queue is a timeout, which is the normal state of a
+		// machine with nothing to build, and is asked again.
+		fetched, err := sub.Fetch(1, nats.Context(ctx))
+		switch {
+		case errors.Is(err, context.Canceled), errors.Is(err, context.DeadlineExceeded):
+			return nil
+		case errors.Is(err, nats.ErrTimeout):
+			continue
+		case err != nil:
+			if sub.IsValid() {
+				// A transient fault in asking — a reconnect, a slow server — is asked past rather
+				// than ending the machine; one that outlasts the ack wait redelivers nothing lost.
+				time.Sleep(time.Second)
+				continue
 			}
+			return fmt.Errorf("the bus stopped delivering build work: %w", err)
+		}
+		for _, msg := range fetched {
 			var request BuildRequest
 			if err := json.Unmarshal(msg.Data, &request); err != nil {
 				// Unreadable: terminated rather than retried, because the next attempt reads the same

internal/link/builds_nats_test.go

@@ -48,7 +48,7 @@ func aBusWithTheBuildRole(t *testing.T) *broker.JetStream {
 		t.Fatal(err)
 	}
 	clean := func() {
-		_ = js.Context().DeleteStream("SEAT_MESH_BUILD_MACHINE")
+		_ = js.Context().DeleteStream("SEAT_NODE_BUILD_AGENT")
 		for _, s := range broker.MeshStreams() {
 			_ = js.Context().PurgeStream(s.Name)
 		}
@@ -158,7 +158,7 @@ func TestNatsABuildIsTakenAndItsOutcomeReachesEverybody(t *testing.T) {
 	// And the work left the queue: a request a machine took and settled must not be given to another.
 	deadline := time.Now().Add(5 * time.Second)
 	for time.Now().Before(deadline) {
-		info, err := js.Context().StreamInfo("SEAT_MESH_BUILD_MACHINE")
+		info, err := js.Context().StreamInfo("SEAT_NODE_BUILD_AGENT")
 		if err == nil && info.State.Msgs == 0 {
 			return
 		}
@@ -177,7 +177,7 @@ func TestNatsABuildWaitsForAMachineRatherThanFailing(t *testing.T) {
 	if _, err := js.Context().Publish(BuildWork(), body); err != nil {
 		t.Fatal(err)
 	}
-	info, err := js.Context().StreamInfo("SEAT_MESH_BUILD_MACHINE")
+	info, err := js.Context().StreamInfo("SEAT_NODE_BUILD_AGENT")
 	if err != nil || info.State.Msgs != 1 {
 		t.Fatalf("the work did not queue: %+v %v", info, err)
 	}
@@ -245,3 +245,83 @@ func TestNatsWorkAMachineDidNotAnswerGoesBackToTheQueue(t *testing.T) {
 func quietLog() *log.Logger { return log.New(io.Discard, "", 0) }
 
 var _ = quietLog
+
+// Two machines holding the role share one queue (novox/hq ADR 0190): three asks, each machine takes
+// one and the third waits until one of them is done; an ask is never handed to a machine that is
+// busy; and a machine that stops mid-ask leaves its ask to the other.
+func TestNatsTwoMachinesShareTheWorkAndNeitherIsHandedMoreThanItCanTake(t *testing.T) {
+	js := aBusWithTheBuildRole(t)
+	ctx, stop := context.WithCancel(context.Background())
+	defer stop()
+
+	for _, id := range []string{"w-1", "w-2", "w-3"} {
+		body, _ := json.Marshal(BuildRequest{ID: id, Repository: "/r"})
+		if _, err := js.Context().Publish(BuildWork(), body); err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	type taken struct{ machine, id string }
+	took := make(chan taken, 8)
+	release := map[string]chan struct{}{"anchor": make(chan struct{}), "laptop": make(chan struct{})}
+	machines := map[string]BuildMachine{}
+	for _, name := range []string{"anchor", "laptop"} {
+		name := name
+		m := MachineOverNATS(js, name)
+		machines[name] = m
+		defer m.Close()
+		go func() {
+			_ = m.Take(ctx, func(ctx context.Context, work Build) {
+				took <- taken{name, work.Request().ID}
+				<-release[name]
+				_ = work.Announce(ctx, BuildResult{ID: work.Request().ID, On: name})
+				_ = work.Done()
+			})
+		}()
+	}
+
+	// Each machine took exactly one, and they are different asks.
+	first := map[string]string{}
+	for i := 0; i < 2; i++ {
+		select {
+		case got := <-took:
+			if _, twice := first[got.machine]; twice {
+				t.Fatalf("%s was handed a second ask while busy with its first", got.machine)
+			}
+			first[got.machine] = got.id
+		case <-time.After(10 * time.Second):
+			t.Fatalf("only %d machine(s) took work; two idle holders should both have", len(first))
+		}
+	}
+	if first["anchor"] == first["laptop"] {
+		t.Fatalf("both machines took %q: the queue is not shared, it is copied", first["anchor"])
+	}
+	// The third waits: nobody is free.
+	select {
+	case got := <-took:
+		t.Fatalf("%s was handed %s while both machines were busy", got.machine, got.id)
+	case <-time.After(2 * time.Second):
+	}
+	// One finishes, and only then is the third taken — by that machine, the one that is free.
+	close(release["anchor"])
+	release["anchor"] = make(chan struct{})
+	select {
+	case got := <-took:
+		if got.machine != "anchor" {
+			t.Fatalf("the third ask went to %s, which is still busy", got.machine)
+		}
+	case <-time.After(10 * time.Second):
+		t.Fatal("the third ask was never taken after a machine became free")
+	}
+	// A machine that stops mid-ask leaves its ask unacknowledged, and the ack wait brings it round
+	// to whoever is left — the path TestNatsWorkAMachineDidNotAnswerGoesBackToTheQueue proves with
+	// an explicit hand-back, because the real wait is a minute. Here: the laptop goes, anchor
+	// finishes, and with nothing queued nothing more is taken by the machine that is left.
+	machines["laptop"].Close()
+	close(release["anchor"])
+	select {
+	case got := <-took:
+		t.Fatalf("%s took %s; the queue should be empty", got.machine, got.id)
+	case <-time.After(2 * time.Second):
+	}
+}