Merge pull request 'The checker runs beside today's check: checks wait while a delivery builds; a group's SDK member merges; a flaky test fixed' (#430) from feat/the-checker into main

internal/broker/checks.go

@@ -46,6 +46,7 @@ func CheckStreams() []Stream {
 			Retention:         RetentionWorkQueue,
 			MaxAge:            7 * 24 * 60 * 60,
 			MaxMsgsPerSubject: 1,
+			Direct:            true,
 			Why: "the checks of pull requests marked ready, taken after a delivery build and before any other check " +
 				"(novox/hq ADR 0406 §3); one subject a pull request, the newest request kept, so a newer one replaces a " +
 				"request still waiting",
@@ -56,6 +57,7 @@ func CheckStreams() []Stream {
 			Retention:         RetentionWorkQueue,
 			MaxAge:            7 * 24 * 60 * 60,
 			MaxMsgsPerSubject: 1,
+			Direct:            true,
 			Why: "every other check, taken last and in the order asked (novox/hq ADR 0406 §3); one subject a pull " +
 				"request, the newest request kept; one left behind for a head already judged is dropped when taken",
 		},
@@ -98,6 +100,13 @@ func holderCheckGrants() []string {
 		"$KV."+CheckRunningBucket+".>",
 		"$JS.API.STREAM.INFO.KV_"+CheckRunningBucket,
 		"$JS.API.DIRECT.GET.KV_"+CheckRunningBucket+".>")
+	// And, to hold checks while a delivery builds (ADR 0483, issues 1687 and 1688): the bucket of the plans building, **read and
+	// never written** (the controller's alone, so no `$KV.` subject of it is granted), and the oldest waiting ask of each check
+	// queue, read by a direct get that takes nothing from the queue and so counts against no delivery of the ask.
+	pub = append(pub, readGrants(PlansBuildingBucket, "")...)
+	for _, w := range CheckWorkers() {
+		pub = append(pub, "$JS.API.DIRECT.GET."+w.Stream, "$JS.API.DIRECT.GET."+w.Stream+".>")
+	}
 	return pub
 }
 

internal/broker/checks_test.go

@@ -40,3 +40,37 @@ func TestTheBuildAgentReadsAndWatchesThePullRequestsKeys(t *testing.T) {
 		}
 	}
 }
+
+// Issue 1687 (ADR 0483 clause 2): a holder of the build seat reads the bucket of the plans building and writes none of it, and
+// reads the oldest waiting ask of each check queue by a direct get, which takes nothing from the queue.
+func TestTheBuildSeatsHoldersReadThePlansBuildingAndWriteNone(t *testing.T) {
+	seat := Seat{Name: TheBuildSeat, Scope: "node", Accepts: []string{"build"}, Emits: []string{"built", "started"}}
+	p := Principal{Kind: KindModule, Node: "one", Module: "build-agent", Holds: []Seat{seat}, PasswordHash: "x"}
+	perms, err := PermissionsFor(p)
+	if err != nil {
+		t.Fatal(err)
+	}
+	stream := "KV_" + PlansBuildingBucket
+	for _, s := range []string{
+		"$JS.API.STREAM.INFO." + stream,
+		"$JS.API.DIRECT.GET." + stream + ".>",
+		"$JS.API.CONSUMER.CREATE." + stream + ".>",
+		"$JS.API.DIRECT.GET.CHECKS_READY",
+		"$JS.API.DIRECT.GET.CHECKS_OTHER",
+	} {
+		has(t, perms.Publish, s)
+	}
+	for _, s := range perms.Publish {
+		if SubjectsOverlap(s, "$KV."+PlansBuildingBucket+".plan-1") {
+			t.Errorf("a holder of the build seat may write %s through %q", PlansBuildingBucket, s)
+		}
+	}
+	// And the controller alone writes it, as the writers table says.
+	for _, row := range WritersTable {
+		for _, s := range row.Subjects {
+			if s == "$KV."+PlansBuildingBucket+".>" && row.Writes(p, "$KV."+PlansBuildingBucket+".plan-1") {
+				t.Errorf("%q lets the build seat's holder write the plans building", row.State)
+			}
+		}
+	}
+}

internal/broker/controller_buckets.go

@@ -63,6 +63,26 @@ var (
 	// repository: the controller reads the ask's warrant from the router and accepts only the operator's Approve of
 	// those heads. Written by the controller alone, from the announcement, when it holds the lease.
 	DeliveryMergesBucket = BucketName(ControllerSeat, "delivery-merges")
+	// PlansBuildingBucket holds one key per delivery plan in its build phase (novox/hq ADR 0483 clause 2, issue 1687): the key
+	// is `<plan id>` (`plan-<digits>`), its value the plan, the time its build phase began and the time of the last refresh.
+	// The controller (lease holder) puts it when the plan's first build is asked, puts it again about every thirty seconds
+	// while the plan builds, and deletes it when the phase ends (every build of the last tier built, or the plan fails, stops
+	// or waits on the operator); the bucket's age, PlansBuildingKeptFor, is how long a key outlives a controller that
+	// stopped. **Read by the build seat's holders and the checker, written by the controller alone**: a holder takes no new
+	// check while any key is there, but for one that waited past ADR 0483's bound. A reader lists the keys and reads none of
+	// their values to decide that (`link.BuildingPlans`).
+	PlansBuildingBucket = BucketName(ControllerSeat, PlansBuilding)
+)
+
+// PlansBuilding is the local name of the bucket of the plans building, as a reader names it in `reads`
+// (`mesh-controller.plans-building`).
+const PlansBuilding = "plans-building"
+
+// PlansBuildingKeptFor is the age of a plan-building key: the controller puts it again at PlansBuildingRefresh, so a
+// controller that stopped leaves checks held at most this long (ADR 0483 clause 2).
+const (
+	PlansBuildingKeptFor = 2 * time.Minute
+	PlansBuildingRefresh = 30 * time.Second
 )
 
 // DeliveryMergesKeptFor is how long a merge mesh-delivery announced is kept: a week, as the events are.
@@ -133,7 +153,7 @@ func ChainBuckets() []string { return []string{BuildsBucket, RolloutsBucket, Ass
 // ControllerMirrors are the local names of the controller's buckets a module may read: the record mirrors and the
 // chain's buckets.
 func ControllerMirrors() []string {
-	return []string{MirrorAssignments, MirrorDeclaredConfig, MirrorPredefined, ChainBuilds, ChainRollouts, ChecksRunning}
+	return []string{MirrorAssignments, MirrorDeclaredConfig, MirrorPredefined, ChainBuilds, ChainRollouts, ChecksRunning, PlansBuilding}
 }
 
 // MirrorsWrittenByAnother are the controller's mirrors whose writer is not the controller, each with its reason: the rule
@@ -204,14 +224,14 @@ func IsControllerBucket(bucket string) bool {
 		bucket == AssignmentsBucket || bucket == DeclaredConfigBucket || bucket == PredefinedBucket ||
 		bucket == MergeWindowBucket || bucket == RolloutSilentBucket || bucket == CheckRunningBucket ||
 		bucket == DeliveryMergesBucket || bucket == BuildsBucket || bucket == RolloutsBucket ||
-		bucket == delivery.HaveBucket
+		bucket == PlansBuildingBucket || bucket == delivery.HaveBucket
 }
 
 // ControllerBuckets are the controller's own buckets, in the order they are asserted.
 func ControllerBuckets() []string {
 	return []string{LeaseBucket, CallsBucket, HandActsBucket, ConditionsBucket, ConditionHistoryBucket, AskedBucket,
 		AssignmentsBucket, DeclaredConfigBucket, PredefinedBucket, MergeWindowBucket, RolloutSilentBucket, CheckRunningBucket,
-		DeliveryMergesBucket, BuildsBucket, RolloutsBucket, delivery.HaveBucket}
+		DeliveryMergesBucket, PlansBuildingBucket, BuildsBucket, RolloutsBucket, delivery.HaveBucket}
 }
 
 // ControllerBucketsAsserter is what raising the controller's buckets needs of a connection.
@@ -354,6 +374,22 @@ func (j *JetStream) EnsureControllerBuckets() error {
 	}); err != nil {
 		return fmt.Errorf("asserting bucket %s: %w", CheckRunningBucket, err)
 	}
+	// **The plans building: keys that age** (ADR 0483 clause 2, issue 1687). One message a key, each refresh a new one, so the
+	// age of the bucket is how long a key outlives a controller that stopped refreshing it.
+	if _, err := js.CreateOrUpdateKeyValue(ctx, jetstream.KeyValueConfig{
+		Bucket: PlansBuildingBucket,
+		Description: "the delivery plans in their build phase (novox/hq ADR 0483): `<plan id>` holds the plan and when its " +
+			"build phase began; put when the plan's first build is asked and about every thirty seconds, deleted when the " +
+			"phase ends, gone within two minutes of a controller that stopped; read by the build seat's holders, which " +
+			"take no new check while a key is there; written by the controller alone",
+		History:      1,
+		TTL:          PlansBuildingKeptFor,
+		MaxValueSize: 4 << 10,
+		MaxBytes:     8 << 20,
+		Storage:      jetstream.FileStorage,
+	}); err != nil {
+		return fmt.Errorf("asserting bucket %s: %w", PlansBuildingBucket, err)
+	}
 	// **The merges mesh-delivery announced: keys that age** (ADR 0420 §7, issue 1025). One message a key; a week.
 	if _, err := js.CreateOrUpdateKeyValue(ctx, jetstream.KeyValueConfig{
 		Bucket: DeliveryMergesBucket,

internal/broker/nats.go

@@ -424,6 +424,11 @@ var DegradableGrants = []DegradableGrant{
 		Reason: "refused: the controller asserts this bucket through the API and writes none of its keys — the build seat's " +
 			"holders do (ADR 0406 §3, issue 974)",
 	},
+	{
+		Subject: "$KV." + PlansBuildingBucket + ".>",
+		Reason: "refused: the controller then says no plan builds, and no check is held while a delivery builds — a check run " +
+			"wrongly costs a delivery a place, a check held wrongly costs a pull request its turn (ADR 0483 clause 2, issue 1687)",
+	},
 	{
 		Subject: "mesh.seat." + ControllerSeat + ".event.sent",
 		Reason: "refused: the controller says a tier sent to a machine only for outsiders, and a send that could not " +

internal/broker/plans_building_live_test.go

@@ -0,0 +1,134 @@
+//go:build integration
+
+package broker
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"github.com/nats-io/nats-server/v2/server"
+	"github.com/nats-io/nats.go"
+	"github.com/nats-io/nats.go/jetstream"
+	"golang.org/x/crypto/bcrypt"
+)
+
+// **A holder of the build seat reads the plans building and the server refuses its write** (novox/hq ADR 0483 clause 2, issue
+// 1687): on a real server whose users are composed as the controller composes them, the controller makes the bucket and puts a
+// plan's key; the build agent's credential lists the keys and gets one, and its put of a key is refused with a permissions
+// violation. It also reads the first waiting ask of a check queue by a direct get, which takes nothing from the queue.
+func TestNatsTheBuildSeatsHoldersReadThePlansBuildingAndTheServerRefusesTheirWrite(t *testing.T) { // 1687
+	hash, err := bcrypt.GenerateFromPassword([]byte("pw"), bcrypt.MinCost)
+	if err != nil {
+		t.Fatal(err)
+	}
+	seat := Seat{Name: TheBuildSeat, Scope: "node", Accepts: []string{"build"}, Emits: []string{"built", "started"}}
+	users, err := Users(Records{Nodes: []string{"alpha"}, Assigned: map[string][]Declared{
+		"alpha": {{Module: RuntimeModule}, {Module: "build-agent", Holds: []Seat{seat}}},
+	}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	for i := range users {
+		users[i].PasswordHash = string(hash)
+	}
+	accounts, err := ComposeAccounts(users)
+	if err != nil {
+		t.Fatal(err)
+	}
+	conf := filepath.Join(t.TempDir(), "bus.conf")
+	if err := os.WriteFile(conf, []byte("listen: 127.0.0.1:-1\njetstream { store_dir: "+
+		`"`+t.TempDir()+`"`+" }\n"+accounts), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	opts, err := server.ProcessConfigFile(conf)
+	if err != nil {
+		t.Fatalf("the composed accounts do not parse: %v", err)
+	}
+	opts.NoLog, opts.NoSigs = true, true
+	s, err := server.NewServer(opts)
+	if err != nil {
+		t.Fatal(err)
+	}
+	go s.Start()
+	if !s.ReadyForConnections(10 * time.Second) {
+		t.Fatal("the bus did not come up")
+	}
+	t.Cleanup(s.Shutdown)
+
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+	connect := func(user string) (*nats.Conn, chan error) {
+		t.Helper()
+		refused := make(chan error, 8)
+		nc, err := nats.Connect(s.ClientURL(), nats.UserInfo(user, "pw"), nats.CustomInboxPrefix("_INBOX."+user),
+			nats.ErrorHandler(func(_ *nats.Conn, _ *nats.Subscription, err error) { refused <- err }))
+		if err != nil {
+			t.Fatalf("%s cannot connect: %v", user, err)
+		}
+		t.Cleanup(nc.Close)
+		return nc, refused
+	}
+	controller, _ := connect(ControllerName)
+	cj := OnConn(controller)
+	if err := AssertMeshStreams(cj); err != nil {
+		t.Fatalf("the controller cannot make the streams: %v", err)
+	}
+	if err := cj.EnsureControllerBuckets(); err != nil {
+		t.Fatalf("the controller cannot make its buckets: %v", err)
+	}
+	cjs, _ := jetstream.New(controller)
+	ckv, err := cjs.KeyValue(ctx, PlansBuildingBucket)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ckv.Put(ctx, "plan-1", []byte(`{"plan":"plan-1"}`)); err != nil {
+		t.Fatalf("the controller cannot put a plan's key: %v", err)
+	}
+
+	holder, refused := connect("alpha.build-agent")
+	hjs, err := jetstream.New(holder)
+	if err != nil {
+		t.Fatal(err)
+	}
+	kv, err := hjs.KeyValue(ctx, PlansBuildingBucket)
+	if err != nil {
+		t.Fatalf("a holder of the build seat cannot bind %s: %v", PlansBuildingBucket, err)
+	}
+	keys, err := kv.Keys(ctx)
+	if err != nil || len(keys) != 1 || keys[0] != "plan-1" {
+		t.Fatalf("the holder lists %v (%v), want plan-1", keys, err)
+	}
+	if e, err := kv.Get(ctx, "plan-1"); err != nil || e == nil {
+		t.Fatalf("the holder cannot get the key: %v", err)
+	}
+	// Its write is refused by the server.
+	if err := holder.Publish("$KV."+PlansBuildingBucket+".plan-2", []byte(`{}`)); err != nil {
+		t.Fatal(err)
+	}
+	_ = holder.Flush()
+	select {
+	case err := <-refused:
+		if !errors.Is(err, nats.ErrPermissionViolation) {
+			t.Errorf("the server said %v, want a permissions violation", err)
+		}
+	case <-time.After(3 * time.Second):
+		t.Error("the server did not refuse the holder's put of a plan's key")
+	}
+	if _, err := ckv.Get(ctx, "plan-2"); !errors.Is(err, jetstream.ErrKeyNotFound) {
+		t.Errorf("the refused put left a key: %v", err)
+	}
+
+	// And the direct get of a check queue's first waiting ask, which is granted.
+	reply, err := holder.RequestWithContext(ctx, "$JS.API.DIRECT.GET."+ChecksOtherStream, []byte(`{"seq":1,"next_by_subj":"`+
+		CheckSubjectPrefix(TheBuildSeat, false)+`>"}`))
+	if err != nil {
+		t.Fatalf("the holder's direct get of %s was not answered: %v", ChecksOtherStream, err)
+	}
+	if reply.Header.Get("Status") != "404" {
+		t.Errorf("an empty queue answered %q, want 404", reply.Header.Get("Status"))
+	}
+}

internal/broker/plans_building_mirror_test.go

@@ -0,0 +1,30 @@
+package broker
+
+import (
+	"slices"
+	"testing"
+)
+
+// Issue 1687 (ADR 0483 clause 2): a module whose manifest reads mesh-controller.plans-building may name it in `reads`, and
+// reads the bucket and writes none; the controller writes it, so it is no mirror written by another.
+func TestAModuleReadingPlansBuildingReadsAndWritesNone(t *testing.T) {
+	if !slices.Contains(ControllerMirrors(), "plans-building") || BucketName(ControllerSeat, PlansBuilding) != PlansBuildingBucket {
+		t.Fatalf("a module cannot name %s in `reads`: %v", PlansBuildingBucket, ControllerMirrors())
+	}
+	if _, other := MirrorsWrittenByAnother[PlansBuilding]; other {
+		t.Error("the controller writes the plans building itself")
+	}
+	perms, err := PermissionsFor(Principal{Kind: KindModule, Node: "one", Module: "mesh-delivery-checker", PasswordHash: "x",
+		Reads: []string{"mesh-controller.plans-building"}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	stream := "KV_" + PlansBuildingBucket
+	has(t, perms.Publish, "$JS.API.STREAM.INFO."+stream)
+	has(t, perms.Publish, "$JS.API.DIRECT.GET."+stream+".>")
+	for _, p := range perms.Publish {
+		if SubjectsOverlap(p, "$KV."+PlansBuildingBucket+".plan-1") {
+			t.Errorf("the reading module may write %s through %q", PlansBuildingBucket, p)
+		}
+	}
+}

internal/broker/testdata/composed.conf

@@ -29,7 +29,7 @@ accounts {
     ]
     users = [
       { user: "controller", password: "$2a$11$cccccccccccccccccccccc", permissions: {
-        publish: { allow: ["$JS.ACK.CONTROL.controller.>", "$JS.ACK.DEAD_LETTER_NOTICES.controller.>", "$JS.ACK.EVENTS.controller-merges.>", "$JS.ACK.EVENTS.controller.>", "$JS.API.>", "$KV.SEAT_MESH_BUILD_MACHINE_cancelled.>", "$KV.SEAT_NODE_BUILD_AGENT_cancelled.>", "$KV.mesh-controller_asked.>", "$KV.mesh-controller_assignments.>", "$KV.mesh-controller_builds.>", "$KV.mesh-controller_calls.>", "$KV.mesh-controller_checks-running.>", "$KV.mesh-controller_condition-history.>", "$KV.mesh-controller_conditions.>", "$KV.mesh-controller_declared-config.>", "$KV.mesh-controller_delivery-merges.>", "$KV.mesh-controller_hand-acts.>", "$KV.mesh-controller_lease.>", "$KV.mesh-controller_merge-window.>", "$KV.mesh-controller_predefined.>", "$KV.mesh-controller_rollout-silent.>", "$KV.mesh-controller_rollouts.>", "$SRV.INFO", "_INBOX.enrol.>", "mesh.again.>", "mesh.assignment.>", "mesh.events.dead.>", "mesh.mod.*.tool.>", "mesh.node.>", "mesh.seat.mesh-build-machine.accept.>", "mesh.seat.mesh-build-machine.tool.>", "mesh.seat.mesh-controller.event.applied", "mesh.seat.mesh-controller.event.assignment-changed", "mesh.seat.mesh-controller.event.built-before", "mesh.seat.mesh-controller.event.checked", "mesh.seat.mesh-controller.event.condition-changed", "mesh.seat.mesh-controller.event.condition-cleared", "mesh.seat.mesh-controller.event.condition-raised", "mesh.seat.mesh-controller.event.declared-config-changed", "mesh.seat.mesh-controller.event.doctor-heartbeat", "mesh.seat.mesh-controller.event.healer-acted", "mesh.seat.mesh-controller.event.plan-moved", "mesh.seat.mesh-controller.event.predefined-changed", "mesh.seat.mesh-controller.event.refused", "mesh.seat.mesh-controller.event.rolled-back", "mesh.seat.mesh-controller.event.secret-replaced", "mesh.seat.mesh-controller.event.sent", "mesh.seat.mesh-controller.tool.plans", "mesh.seat.mesh-controller.tool.push", "mesh.seat.mesh-controller.tool.upgrade", "mesh.seat.mesh-delivery.tool.close", "mesh.seat.mesh-delivery.tool.groups", "mesh.seat.mesh-delivery.tool.release", "mesh.seat.mesh-delivery.tool.stalled", "mesh.seat.mesh-delivery.tool.stop", "mesh.seat.mesh-delivery.tool.times", "mesh.seat.node-backup.tool.backed-up.*", "mesh.seat.node-backup.tool.now.*", "mesh.seat.node-build-agent.accept.>", "mesh.seat.node-build-agent.check.>", "mesh.seat.node-build-agent.tool.>", "mesh.seat.node-intrusion-prevention.tool.banned.*", "mesh.seat.node-launcher.tool.secret.*", "mesh.seat.node-service-manager.tool.restart.*", "mesh.seat.operator-channel.tool.tell"] }
+        publish: { allow: ["$JS.ACK.CONTROL.controller.>", "$JS.ACK.DEAD_LETTER_NOTICES.controller.>", "$JS.ACK.EVENTS.controller-merges.>", "$JS.ACK.EVENTS.controller.>", "$JS.API.>", "$KV.SEAT_MESH_BUILD_MACHINE_cancelled.>", "$KV.SEAT_NODE_BUILD_AGENT_cancelled.>", "$KV.mesh-controller_asked.>", "$KV.mesh-controller_assignments.>", "$KV.mesh-controller_builds.>", "$KV.mesh-controller_calls.>", "$KV.mesh-controller_checks-running.>", "$KV.mesh-controller_condition-history.>", "$KV.mesh-controller_conditions.>", "$KV.mesh-controller_declared-config.>", "$KV.mesh-controller_delivery-merges.>", "$KV.mesh-controller_hand-acts.>", "$KV.mesh-controller_lease.>", "$KV.mesh-controller_merge-window.>", "$KV.mesh-controller_plans-building.>", "$KV.mesh-controller_predefined.>", "$KV.mesh-controller_rollout-silent.>", "$KV.mesh-controller_rollouts.>", "$SRV.INFO", "_INBOX.enrol.>", "mesh.again.>", "mesh.assignment.>", "mesh.events.dead.>", "mesh.mod.*.tool.>", "mesh.node.>", "mesh.seat.mesh-build-machine.accept.>", "mesh.seat.mesh-build-machine.tool.>", "mesh.seat.mesh-controller.event.applied", "mesh.seat.mesh-controller.event.assignment-changed", "mesh.seat.mesh-controller.event.built-before", "mesh.seat.mesh-controller.event.checked", "mesh.seat.mesh-controller.event.condition-changed", "mesh.seat.mesh-controller.event.condition-cleared", "mesh.seat.mesh-controller.event.condition-raised", "mesh.seat.mesh-controller.event.declared-config-changed", "mesh.seat.mesh-controller.event.doctor-heartbeat", "mesh.seat.mesh-controller.event.healer-acted", "mesh.seat.mesh-controller.event.plan-moved", "mesh.seat.mesh-controller.event.predefined-changed", "mesh.seat.mesh-controller.event.refused", "mesh.seat.mesh-controller.event.rolled-back", "mesh.seat.mesh-controller.event.secret-replaced", "mesh.seat.mesh-controller.event.sent", "mesh.seat.mesh-controller.tool.plans", "mesh.seat.mesh-controller.tool.push", "mesh.seat.mesh-controller.tool.upgrade", "mesh.seat.mesh-delivery.tool.close", "mesh.seat.mesh-delivery.tool.groups", "mesh.seat.mesh-delivery.tool.release", "mesh.seat.mesh-delivery.tool.stalled", "mesh.seat.mesh-delivery.tool.stop", "mesh.seat.mesh-delivery.tool.times", "mesh.seat.node-backup.tool.backed-up.*", "mesh.seat.node-backup.tool.now.*", "mesh.seat.node-build-agent.accept.>", "mesh.seat.node-build-agent.check.>", "mesh.seat.node-build-agent.tool.>", "mesh.seat.node-intrusion-prevention.tool.banned.*", "mesh.seat.node-launcher.tool.secret.*", "mesh.seat.node-service-manager.tool.restart.*", "mesh.seat.operator-channel.tool.tell"] }
         subscribe: { allow: ["$JS.API.>", "$JS.EVENT.ADVISORY.CONSUMER.DELETED.>", "$JS.EVENT.ADVISORY.CONSUMER.MAX_DELIVERIES.>", "$SRV.INFO", "$SRV.INFO.mesh-controller", "$SRV.INFO.mesh-controller.>", "$SRV.PING", "$SRV.PING.mesh-controller", "$SRV.PING.mesh-controller.>", "$SRV.STATS", "$SRV.STATS.mesh-controller", "$SRV.STATS.mesh-controller.>", "_DELIVER.controller", "_DELIVER.controller.>", "_INBOX.controller.>", "mesh.control.>", "mesh.mod.*.event.provisioner.failing", "mesh.mod.*.event.provisioner.recovered", "mesh.mod.*.event.provisioner.retirement", "mesh.mod.gitea.event.pull.merged", "mesh.mod.gitea.event.pull.updated", "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.>", "mesh.seat.node-build-agent.event.built", "mesh.seat.operator-channel.event.decided.mesh-controller"] }
         allow_responses: { max: 1, ttl: "1m" }
       } }

internal/broker/writers.go

@@ -177,6 +177,10 @@ var WritersTable = []WriterRow{
 		KeptIn: "key-value " + CheckRunningBucket + ", aged out when not renewed", Others: "the build seat's holders read it to drop a duplicate request",
 		Subjects: kvOf(CheckRunningBucket), Writes: holdsTheBuildSeat,
 		Shared: "every machine holding the build seat puts the key of the check it runs and deletes it when the check ends"},
+	{State: "the plans-building mirror", Writer: "controller (lease holder), at the first build's ask and at each refresh",
+		KeptIn:   "key-value " + PlansBuildingBucket + ", aged out when not refreshed",
+		Others:   "the build seat's holders read it and take no new check while a key is there (ADR 0483)",
+		Subjects: kvOf(PlansBuildingBucket), Writes: isController},
 	{State: "a merge announced", Writer: "one announcer per forge (the hook, or the poll when the hook is absent — never both)",
 		KeptIn: "the bus", Others: "—", Subjects: []string{"mesh.mod.*.event.pull.merged"}, Writes: ownModule},
 	// A pull request's head, before it merges (novox/hq to-be 45 §9): the same one announcer.

internal/broker/writers_test.go

@@ -29,6 +29,8 @@ var designRows = []string{
 	"builds and their outcomes",
 	"a pull request's check asked",
 	"a pull request's check running",
+	// The plans in their build phase, a bucket the build seat's holders read (ADR 0483 clause 2, issue 1687).
+	"the plans-building mirror",
 	"a merge announced",
 	"a pull request's head announced",
 	"a pull request's merge check",

internal/link/build.go

@@ -292,6 +292,10 @@ type GroupMember struct {
 	// Base is the branch it merges into: the build seat merges the head into that trunk's head before the
 	// group is composed (novox/hq ADR 0323 §6). A build seat from before it composes the head alone.
 	Base string `json:"base,omitempty"`
+	// Trunk is the base branch's head the forge announced with the head (novox/hq issue 1106): the trunk commit the
+	// composed verdict names for this member when the build seat named none, so a member whose trunk did not move is
+	// never merged-refused for a verdict that names no trunk.
+	Trunk string `json:"trunk,omitempty"`
 }
 
 // Who judges a merge check's gate.

internal/link/builds_nats.go

@@ -6,6 +6,7 @@ import (
 	"errors"
 	"fmt"
 	"os"
+	"sync"
 	"time"
 
 	"github.com/nats-io/nats.go"
@@ -164,11 +165,21 @@ type natsMachine struct {
 	opts MachineOptions
 	// duplicate stands in for runningElsewhere in tests; nil is the real one.
 	duplicate func(context.Context, BuildRequest) string
+	// gate is what the loop asks of the bus to hold checks while a delivery builds; nil is the bus's (a test sets a fake).
+	gate checkGate
+	// said is the faults of the hold already said, letKey whether this holder created the let-through key for the check it
+	// took last (ADR 0483 clause 3).
+	said   map[string]bool
+	letKey bool
+	// mu guards the subscriptions of subs, bound and let go of by the taking loop and closed by Close.
+	mu sync.Mutex
 }
 
 // takeQueue is one queue a machine pulls from.
 type takeQueue struct {
 	sub *nats.Subscription
+	// pull is what the queue is fetched through: the subscription once bound (a test sets a fake).
+	pull puller
 	// kind is what its asks are: a build, a ready check, or another check.
 	kind string
 	// worker is the consumer bound; bound is false until a bind succeeded, for a queue the mesh may not have made yet.
@@ -176,6 +187,11 @@ type takeQueue struct {
 	bound  bool
 }
 
+// puller is the fetch of a queue.
+type puller interface {
+	Fetch(batch int, opts ...nats.PullOpt) ([]*nats.Msg, error)
+}
+
 // The kinds of queue a machine takes from, in the order it takes from them.
 const (
 	queueBuilds = "builds"
@@ -197,6 +213,9 @@ type MachineOptions struct {
 	// the request in hand first and reads its drain after, so an ask of `busy` racing the fetch sees one or
 	// the other, never neither. Nil claims every one.
 	Claim func() bool
+	// LetThroughAfter is how long a check waits while a delivery builds before it is let through; zero is ADR 0483's thirty
+	// minutes. Set by a test.
+	LetThroughAfter time.Duration
 }
 
 // MachineOverNATS takes build work from the current build role.
@@ -224,6 +243,8 @@ const (
 )
 
 func (m *natsMachine) Close() {
+	m.mu.Lock()
+	defer m.mu.Unlock()
 	for _, q := range m.subs {
 		if q.sub != nil {
 			_ = q.sub.Unsubscribe()
@@ -246,6 +267,9 @@ func (m *natsMachine) Close() {
 // **In this order** (novox/hq ADR 0406 §3): a delivery build, then a check of a pull request marked ready, then any
 // other check, each queue tried without waiting; when all three are empty the first is waited on for a second at most
 // and the three are tried again, so a delivery build is never much more than a second behind and nothing spins.
+//
+// **Unless a delivery builds** (novox/hq ADR 0483): while the controller's bucket of the plans building holds a key, no check queue
+// is fetched from, but for the one check that waited LetThroughAfter, which goes first (see next).
 func (m *natsMachine) Take(ctx context.Context, do func(context.Context, Build)) error {
 	worker, found := broker.HolderConsumerFor(m.on, "build-agent",
 		broker.DeclaredSeat{Name: m.seat, Accepts: []string{"build"}})
@@ -290,6 +314,19 @@ func (m *natsMachine) Take(ctx context.Context, do func(context.Context, Build))
 			}
 			continue
 		}
+		// The seat's own queue is bound again when a running ask made the machine let go of it (quiesce): a machine that cannot
+		// is not taking, which is said and tried again.
+		if !m.subs[0].bound {
+			if err := m.bind(m.subs[0]); err != nil {
+				fmt.Fprintf(os.Stderr, "this machine cannot bind %s again after a build, and tries again: %v\n", m.seat, err)
+				select {
+				case <-ctx.Done():
+					return nil
+				case <-time.After(pausedPoll):
+				}
+				continue
+			}
+		}
 		if time.Now().After(bindAgain) {
 			for _, q := range m.subs[1:] {
 				if !q.bound {
@@ -314,7 +351,7 @@ func (m *natsMachine) Take(ctx context.Context, do func(context.Context, Build))
 			// say why (2026-10-03, the first build agents). Asked again.
 			continue
 		default:
-			if m.subs[0].sub.IsValid() {
+			if m.subs[0].sub != nil && m.subs[0].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)
@@ -322,56 +359,95 @@ func (m *natsMachine) Take(ctx context.Context, do func(context.Context, Build))
 			}
 			return fmt.Errorf("the bus stopped delivering build work: %w", err)
 		}
-		// **Paused while the pull was answered** (ADR 0219): "takes no new build" holds even for
-		// the one that arrived in that moment. Handed back at once for another holder — counted as
-		// a delivery, which a pause landing exactly then costs and nothing else does.
-		if m.opts.Paused != nil && m.opts.Paused() {
-			_ = msg.Nak()
+		m.handOver(ctx, msg, q, do)
+		// And the queues it let go of are bound again at once, not after the half minute.
+		bindAgain = time.Time{}
+	}
+}
+
+// handOver settles one fetched ask: dropped, handed back or built. **A check let through keeps its key for as long as it
+// stays here** (ADR 0483 clause 3): refreshed while it runs and deleted when it ends, however it ends.
+func (m *natsMachine) handOver(ctx context.Context, msg *nats.Msg, q *takeQueue, do func(context.Context, Build)) {
+	if m.letKey {
+		defer m.keepLetThrough(ctx, heartbeatWhileWorking)()
+	}
+	// **Paused while the pull was answered** (ADR 0219): "takes no new build" holds even for
+	// the one that arrived in that moment. Handed back at once for another holder — counted as
+	// a delivery, which a pause landing exactly then costs and nothing else does.
+	if m.opts.Paused != nil && m.opts.Paused() {
+		_ = msg.Nak()
+		return
+	}
+	var request BuildRequest
+	if err := json.Unmarshal(msg.Data, &request); err != nil {
+		// Unreadable: terminated rather than retried, because the next attempt reads the same
+		// bytes. Nobody waiting hears an answer, which is right — there was no request.
+		_ = msg.Term()
+		return
+	}
+	// **Cancelled in the moment it was fetched** (novox/hq ADR 0219): the controller wrote the
+	// id into the seat's cancelled set before deleting the ask, so one taken in between is
+	// ended here, terminated rather than redelivered, and its outcome said as failed — never
+	// built. A set that cannot be read is said and the ask built: a cancel is a person's
+	// exception, and a holder that refused every build while its grant was missing would
+	// stop the mesh building for it.
+	build := &natsBuild{request: request, msg: msg, on: m.on, js: m.js, seat: m.seat, delivery: deliveryOf(msg)}
+	if cancelled, err := IsCancelled(m.js.Conn(), m.seat, request.ID); err != nil {
+		fmt.Fprintf(os.Stderr, "could not read whether %s was cancelled, so it is built: %v\n", request.ID, err)
+	} else if cancelled {
+		endCancelled(ctx, build)
+		return
+	}
+	// **A check left behind is dropped when it is taken** (novox/hq ADR 0406 §3): a request on the third queue for a
+	// head and a trunk commit its pull request's key already says were judged is acknowledged and not run.
+	if q.kind == queueOther {
+		if why := m.judgedAlready(ctx, request); why != "" {
+			fmt.Fprintf(os.Stderr, "%s was left behind in the check queue and is dropped: %s\n", request.ID, why)
+			_ = msg.Ack()
+			return
+		}
+	}
+	if !m.admit(ctx, msg, request) {
+		return
+	}
+	// A build outlives the acknowledgement window many times over; said while it runs,
+	// as the controller says it for its own long handlers, so the server neither hands
+	// the ask to a second machine nor counts the wait against its deliveries.
+	working := make(chan struct{})
+	go stillWorking(msg, working, heartbeatWhileWorking)
+	// And a check says it is running, to the other builders (ADR 0406 §3): its key put now and at each heartbeat,
+	// deleted when it ends, whatever the outcome.
+	stopRunning := keepRunning(ctx, m.js.Conn(), request, m.on, heartbeatWhileWorking)
+	// **And holds nothing else while it runs** (issue 1747): see quiesce.
+	m.quiesce(q)
+	do(ctx, build)
+	stopRunning()
+	close(working)
+	// **The queue it came from too, once the ask is settled**: a fetch that returned a late message leaves its pull request
+	// open, and the server can deliver one more into the subscription (issue 1747). Take binds every queue again first.
+	m.quiesce(nil)
+}
+
+// quiesce lets go of every queue but the one the running ask came from, for as long as it runs. **A builder never holds a
+// message it does not start** (issue 1747). A message the server sends after a fetch gave up, within the few milliseconds its
+// pull request still lived, lands in that queue's subscription and stays there until the next fetch of it: held, never
+// acknowledged as in progress, for as long as the builder runs its current ask, five or ten minutes for a check, while the
+// holders that could build it are all busy and a delivery waits behind it. The legacy pull subscription cannot be read
+// without sending a new pull request, which could take one more, so it cannot be searched for such a message. It is dropped
+// instead: the subscription is closed, a message in it or on its way is lost to this builder and the server hands it out again
+// when its worker's wait for an acknowledgement ends (a minute), to any holder that is free, and the queue is bound again
+// before the next fetch (Take). The queue the running ask came from keeps its subscription, which its acknowledgements use.
+func (m *natsMachine) quiesce(keep *takeQueue) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	for _, q := range m.subs {
+		if q == keep || !q.bound {
 			continue
 		}
-		var request BuildRequest
-		if err := json.Unmarshal(msg.Data, &request); err != nil {
-			// Unreadable: terminated rather than retried, because the next attempt reads the same
-			// bytes. Nobody waiting hears an answer, which is right — there was no request.
-			_ = msg.Term()
-			continue
+		if q.sub != nil {
+			_ = q.sub.Unsubscribe()
 		}
-		// **Cancelled in the moment it was fetched** (novox/hq ADR 0219): the controller wrote the
-		// id into the seat's cancelled set before deleting the ask, so one taken in between is
-		// ended here, terminated rather than redelivered, and its outcome said as failed — never
-		// built. A set that cannot be read is said and the ask built: a cancel is a person's
-		// exception, and a holder that refused every build while its grant was missing would
-		// stop the mesh building for it.
-		build := &natsBuild{request: request, msg: msg, on: m.on, js: m.js, seat: m.seat, delivery: deliveryOf(msg)}
-		if cancelled, err := IsCancelled(m.js.Conn(), m.seat, request.ID); err != nil {
-			fmt.Fprintf(os.Stderr, "could not read whether %s was cancelled, so it is built: %v\n", request.ID, err)
-		} else if cancelled {
-			endCancelled(ctx, build)
-			continue
-		}
-		// **A check left behind is dropped when it is taken** (novox/hq ADR 0406 §3): a request on the third queue for a
-		// head and a trunk commit its pull request's key already says were judged is acknowledged and not run.
-		if q.kind == queueOther {
-			if why := m.judgedAlready(ctx, request); why != "" {
-				fmt.Fprintf(os.Stderr, "%s was left behind in the check queue and is dropped: %s\n", request.ID, why)
-				_ = msg.Ack()
-				continue
-			}
-		}
-		if !m.admit(ctx, msg, request) {
-			continue
-		}
-		// A build outlives the acknowledgement window many times over; said while it runs,
-		// as the controller says it for its own long handlers, so the server neither hands
-		// the ask to a second machine nor counts the wait against its deliveries.
-		working := make(chan struct{})
-		go stillWorking(msg, working, heartbeatWhileWorking)
-		// And a check says it is running, to the other builders (ADR 0406 §3): its key put now and at each heartbeat,
-		// deleted when it ends, whatever the outcome.
-		stopRunning := keepRunning(ctx, m.js.Conn(), request, m.on, heartbeatWhileWorking)
-		do(ctx, build)
-		stopRunning()
-		close(working)
+		q.sub, q.pull, q.bound = nil, nil, false
 	}
 }
 
@@ -438,25 +514,39 @@ func duplicateOf(running CheckRunning, found bool, r BuildRequest, me string) st
 // 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.
 func (m *natsMachine) bind(q *takeQueue) error {
+	m.mu.Lock()
+	defer m.mu.Unlock()
 	sub, err := m.js.Context().PullSubscribe(q.worker.Filters[0], q.worker.Name,
 		nats.Bind(q.worker.Stream, q.worker.Name), nats.ManualAck())
 	if err != nil {
 		return err
 	}
-	q.sub, q.bound = sub, true
+	q.sub, q.pull, q.bound = sub, sub, true
 	return nil
 }
 
 // next takes the next request in the order of the queues: each tried for a moment, and when all are empty the first waited
 // on for a second. Nothing, with no error, when there was nothing.
 func (m *natsMachine) next(ctx context.Context) (*nats.Msg, *takeQueue, error) {
+	// **While a delivery builds, only delivery builds are taken, but for one check that waited** (novox/hq ADR 0483): the
+	// bucket of the plans building is read before either check queue is fetched from, and a key there holds both. The check
+	// that waited LetThroughAfter goes first, ahead of the builds that wait; the rest of the holders take builds only.
+	if m.seat == broker.TheBuildSeat && len(m.subs) > 1 {
+		if plans := m.building(ctx); len(plans) > 0 {
+			if msg, q := m.letOneThrough(ctx); msg != nil {
+				fmt.Fprintf(os.Stderr, "a check that waited %s is let through while %s builds\n", m.letThroughAfter(), whyHeld(plans))
+				return msg, q, nil
+			}
+			return m.nextBuild(ctx)
+		}
+	}
 	var last error
 	for _, q := range m.subs {
 		if !q.bound {
 			continue
 		}
 		trying, end := context.WithTimeout(ctx, checkTakeWait)
-		fetched, err := q.sub.Fetch(1, nats.Context(trying))
+		fetched, err := q.pull.Fetch(1, nats.Context(trying))
 		end()
 		if len(fetched) > 0 {
 			return fetched[0], q, nil
@@ -472,9 +562,18 @@ func (m *natsMachine) next(ctx context.Context) (*nats.Msg, *takeQueue, error) {
 	if last != nil {
 		return nil, nil, last
 	}
+	return m.nextBuild(ctx)
+}
+
+// nextBuild waits for a second at most for a delivery build, the seat's own queue.
+func (m *natsMachine) nextBuild(ctx context.Context) (*nats.Msg, *takeQueue, error) {
+	if !m.subs[0].bound {
+		// Let go of while an ask ran and not bound again yet (Take binds it before it asks): nothing to take from.
+		return nil, nil, nil
+	}
 	waiting, end := context.WithTimeout(ctx, takeIdleWait)
 	defer end()
-	fetched, err := m.subs[0].sub.Fetch(1, nats.Context(waiting))
+	fetched, err := m.subs[0].pull.Fetch(1, nats.Context(waiting))
 	if len(fetched) > 0 {
 		return fetched[0], m.subs[0], nil
 	}
@@ -637,3 +736,16 @@ func endCancelled(ctx context.Context, b *natsBuild) {
 	}
 	_ = b.msg.Term()
 }
+
+// boundKinds are the kinds of queue this machine holds a subscription to now.
+func (m *natsMachine) boundKinds() []string {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	var out []string
+	for _, q := range m.subs {
+		if q.bound {
+			out = append(out, q.kind)
+		}
+	}
+	return out
+}

internal/link/checks_nats.go

@@ -287,12 +287,16 @@ type CheckRunning struct {
 }
 
 // RunningKeyOf is the key a check is said running under: its pull request's tokens, or `group.<name>` for a group's
-// composed check; none for a try, which no other request replaces.
+// composed check; `try.<id>` for a try, which no other request replaces.
 func RunningKeyOf(r BuildRequest) (string, bool) {
 	c := r.Check
 	switch {
-	case c == nil, c.Try:
+	case c == nil:
 		return "", false
+	case c.Try:
+		// A try says it runs under its own id, which no other request shares: so a reader of the bucket can tell a try that runs
+		// from a check left waiting for a redelivery (ADR 0483 clause 3).
+		return "try." + tokenOf(r.ID), true
 	case c.Group != "":
 		return "group." + tokenOf(c.Group), true
 	case c.Number <= 0:

internal/link/checks_nats_integration_test.go

@@ -6,7 +6,6 @@ import (
 	"context"
 	"encoding/json"
 	"reflect"
-	"strings"
 	"testing"
 	"time"
 
@@ -595,21 +594,12 @@ func TestABusyMachineTakesNoWaitingAskWhateverItsPullsHold(t *testing.T) {
 	case <-time.After(10 * time.Second):
 		t.Fatal("builder a took nothing")
 	}
-	var other *takeQueue
-	for _, q := range a.subs {
-		if q.kind == queueOther && q.bound {
-			other = q
-		}
-	}
-	if other == nil {
-		t.Fatal("builder a has no pull on the other checks")
-	}
-	// The queue is empty, so this request is answered at its expiry with a status message that lands in the pull.
-	reply := strings.TrimSuffix(other.sub.Subject, "*") + "late"
-	if err := js.Conn().PublishRequest("$JS.API.CONSUMER.MSG.NEXT."+other.worker.Stream+"."+other.worker.Name, reply,
-		[]byte(`{"batch":1,"expires":1000000}`)); err != nil {
-		t.Fatal(err)
+	// **A busy machine holds no pull at all** (issues 1649 and 1747): the queues the running ask did not come from are let go of,
+	// so no message, a status message or a late delivery, can sit in one while it runs.
+	if kinds := a.boundKinds(); len(kinds) != 1 || kinds[0] != queueBuilds {
+		t.Fatalf("builder a holds %v while it builds; want only the queue its build came from", kinds)
 	}
+	other := &takeQueue{kind: queueOther, worker: broker.CheckWorkers()[1]}
 	time.Sleep(500 * time.Millisecond)
 	for i, id := range []string{"wait-1", "wait-2"} {
 		if err := asks.Ask(ctx, aPullCheck(id, 20+i, "head-x-000000000", "trunk-1-000000", false)); err != nil {

internal/link/checks_wait_integration_test.go

@@ -0,0 +1,479 @@
+//go:build integration
+
+package link
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"github.com/nats-io/nats.go"
+	"reflect"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/nats-io/nats.go/jetstream"
+
+	"github.com/novox/mesh-controller/internal/broker"
+)
+
+// Checks wait while a delivery builds (novox/hq ADR 0483, issues 1652, 1687, 1688), against a real bus.
+
+// aBusWithTheBuckets is the bus with the build role and the controller's buckets, the bucket of the checks running with an age
+// of two seconds so a key a holder stopped refreshing is seen to expire.
+func aBusWithTheBuckets(t *testing.T) (*broker.JetStream, jetstream.KeyValue) {
+	t.Helper()
+	js := aBusWithTheBuildRole(t)
+	api, err := jetstream.New(js.Conn())
+	if err != nil {
+		t.Fatal(err)
+	}
+	ctx := t.Context()
+	if _, err := api.CreateKeyValue(ctx, jetstream.KeyValueConfig{Bucket: broker.CheckRunningBucket, History: 1, TTL: 2 * time.Second}); err != nil {
+		t.Fatal(err)
+	}
+	plans, err := api.CreateKeyValue(ctx, jetstream.KeyValueConfig{Bucket: broker.PlansBuildingBucket, History: 1, TTL: broker.PlansBuildingKeptFor})
+	if err != nil {
+		t.Fatal(err)
+	}
+	return js, plans
+}
+
+func aCheckAskedAgo(ago time.Duration, number int, ready bool) BuildRequest {
+	r := aPullCheck(fmt.Sprintf("build-%d", time.Now().Add(-ago).UnixNano()), number, fmt.Sprintf("head-%d-0000000", number), "trunk-0-000000", ready)
+	return r
+}
+
+// 1652/AC6: a plan's builds asked in two tiers with a check waiting, the check is handed out only after the last tier's builds,
+// and not in the gap between the tiers.
+func TestNatsACheckIsHandedOutOnlyAfterTheLastTierOfAPlan(t *testing.T) {
+	js, plans := aBusWithTheBuckets(t)
+	ctx, stop := context.WithCancel(t.Context())
+	defer stop()
+	asks := BuildsOn(js, TheBuildMachine)
+
+	if _, err := plans.Put(ctx, "plan-1", []byte(`{"plan":"plan-1"}`)); err != nil {
+		t.Fatal(err)
+	}
+	if err := asks.Ask(ctx, aCheckAskedAgo(time.Minute, 7, true)); err != nil {
+		t.Fatal(err)
+	}
+	if err := asks.Ask(ctx, BuildRequest{ID: "tier-1-a", Repository: "/a"}); err != nil {
+		t.Fatal(err)
+	}
+	took := make(chan string, 16)
+	machine := MachineOverNATS(js, "anchor")
+	defer machine.Close()
+	go func() {
+		_ = machine.Take(ctx, func(ctx context.Context, work Build) {
+			took <- work.Request().ID
+			_ = work.Announce(ctx, BuildResult{ID: work.Request().ID, On: "anchor", Failed: "no"})
+			_ = work.Done()
+		})
+	}()
+	expect := func(want string) {
+		t.Helper()
+		select {
+		case got := <-took:
+			if got != want {
+				t.Fatalf("taken %s, want %s", got, want)
+			}
+		case <-time.After(15 * time.Second):
+			t.Fatalf("%s was not taken", want)
+		}
+	}
+	expect("tier-1-a")
+	// The gap between two tiers: the delivery's queue is empty, the plan's key is there, and the check is not taken.
+	select {
+	case got := <-took:
+		t.Fatalf("%s was taken in the gap between two tiers", got)
+	case <-time.After(3 * time.Second):
+	}
+	if err := asks.Ask(ctx, BuildRequest{ID: "tier-2-a", Repository: "/b"}); err != nil {
+		t.Fatal(err)
+	}
+	expect("tier-2-a")
+	select {
+	case got := <-took:
+		t.Fatalf("%s was taken before the plan's last build was built and its key deleted", got)
+	case <-time.After(2 * time.Second):
+	}
+	// The phase ends: the key is deleted, and the check is handed out.
+	if err := plans.Delete(ctx, "plan-1"); err != nil {
+		t.Fatal(err)
+	}
+	select {
+	case got := <-took:
+		if got == "tier-1-a" || got == "tier-2-a" {
+			t.Fatalf("a build was taken twice: %s", got)
+		}
+	case <-time.After(15 * time.Second):
+		t.Fatal("the check was not handed out once the plan's key was deleted")
+	}
+}
+
+// 1688: two holders racing for checks that waited past the bound: exactly one is let through at a time, by a create that fails
+// while the key is there, and the key is deleted when its check ends, after which the next oldest goes.
+func TestNatsTwoHoldersRacingForOldChecksLetExactlyOneThrough(t *testing.T) {
+	js, plans := aBusWithTheBuckets(t)
+	ctx, stop := context.WithCancel(t.Context())
+	defer stop()
+	// The buckets' age is two seconds here, so the keys are refreshed well within it.
+	was := heartbeatWhileWorking
+	heartbeatWhileWorking = 400 * time.Millisecond
+	t.Cleanup(func() { heartbeatWhileWorking = was })
+	asks := BuildsOn(js, TheBuildMachine)
+	if _, err := plans.Put(ctx, "plan-1", []byte(`{"plan":"plan-1"}`)); err != nil {
+		t.Fatal(err)
+	}
+	for i, ago := range []time.Duration{50 * time.Minute, 40 * time.Minute, 35 * time.Minute} {
+		if err := asks.Ask(ctx, aCheckAskedAgo(ago, i+1, false)); err != nil {
+			t.Fatal(err)
+		}
+	}
+	var mu sync.Mutex
+	running, most := 0, 0
+	started := make(chan string, 8)
+	release := make(chan struct{})
+	do := func(ctx context.Context, work Build) {
+		mu.Lock()
+		running++
+		most = max(most, running)
+		mu.Unlock()
+		started <- work.Request().Check.Repo + fmt.Sprint(work.Request().Check.Number)
+		<-release
+		mu.Lock()
+		running--
+		mu.Unlock()
+		_ = work.Announce(ctx, BuildResult{ID: work.Request().ID, On: "x", Failed: "no"})
+		_ = work.Done()
+	}
+	for _, on := range []string{"ace", "shanks"} {
+		machine := MachineOverNATS(js, on)
+		defer machine.Close()
+		go func() { _ = machine.Take(ctx, do) }()
+	}
+	var order []string
+	for range 3 {
+		select {
+		case id := <-started:
+			order = append(order, id)
+		case <-time.After(20 * time.Second):
+			t.Fatalf("started so far: %v", order)
+		}
+		// While it runs, no second check starts and the key is there.
+		select {
+		case id := <-started:
+			t.Fatalf("%s started beside a check let through", id)
+		case <-time.After(2500 * time.Millisecond):
+		}
+		kv := runningKV(t, js)
+		if _, err := kv.Get(ctx, LetThroughKey); err != nil {
+			t.Fatalf("the let-through key is not there while the check runs: %v", err)
+		}
+		release <- struct{}{}
+		// It is deleted when the check ends: the next check let through creates the key again at once, and a create fails
+		// while the key is there, so its start (the next round) is the proof; after the last one the key stays gone.
+		if len(order) == 3 {
+			deadline := time.Now().Add(10 * time.Second)
+			for {
+				if _, err := kv.Get(ctx, LetThroughKey); err != nil {
+					break
+				}
+				if time.Now().After(deadline) {
+					t.Fatal("the let-through key was not deleted when the last check ended")
+				}
+				time.Sleep(50 * time.Millisecond)
+			}
+		}
+	}
+	if want := []string{"a1", "a2", "a3"}; !reflect.DeepEqual(order, want) {
+		t.Fatalf("checks went %v, want the oldest first %v", order, want)
+	}
+	mu.Lock()
+	defer mu.Unlock()
+	if most != 1 {
+		t.Fatalf("%d checks ran at once while a delivery built", most)
+	}
+}
+
+func runningKV(t *testing.T, js *broker.JetStream) jetstream.KeyValue {
+	t.Helper()
+	api, err := jetstream.New(js.Conn())
+	if err != nil {
+		t.Fatal(err)
+	}
+	kv, err := api.KeyValue(t.Context(), broker.CheckRunningBucket)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return kv
+}
+
+// 1688: the create of the let-through key succeeds for exactly one of holders asking at once, and a key its holder stopped
+// refreshing expires, after which the next create succeeds.
+func TestNatsOneCreateOfTheLetThroughKeySucceedsAndAStoppedHoldersKeyExpires(t *testing.T) {
+	js, _ := aBusWithTheBuckets(t)
+	gate := busGate{conn: js.Conn()}
+	ctx := t.Context()
+	var wg sync.WaitGroup
+	var mu sync.Mutex
+	won := 0
+	for i := range 8 {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			ok, err := gate.TakeLetThrough(ctx, fmt.Sprint("holder-", i))
+			if err != nil {
+				t.Error(err)
+				return
+			}
+			if ok {
+				mu.Lock()
+				won++
+				mu.Unlock()
+			}
+		}()
+	}
+	wg.Wait()
+	if won != 1 {
+		t.Fatalf("%d of 8 holders created the let-through key, want exactly one", won)
+	}
+	// The holder stopped: nothing refreshes the key, and the bucket's age (two seconds here) removes it.
+	deadline := time.Now().Add(15 * time.Second)
+	for {
+		ok, err := gate.TakeLetThrough(ctx, "next")
+		if err != nil {
+			t.Fatal(err)
+		}
+		if ok {
+			break
+		}
+		if time.Now().After(deadline) {
+			t.Fatal("the key of a stopped holder did not expire")
+		}
+		time.Sleep(250 * time.Millisecond)
+	}
+	// And a refreshed key does not expire.
+	for range 6 {
+		if err := gate.KeepLetThrough(ctx, "next"); err != nil {
+			t.Fatal(err)
+		}
+		time.Sleep(700 * time.Millisecond)
+	}
+	if ok, _ := gate.TakeLetThrough(ctx, "other"); ok {
+		t.Fatal("a refreshed key expired")
+	}
+	var v LetThrough
+	e, err := runningKV(t, js).Get(ctx, LetThroughKey)
+	if err != nil || json.Unmarshal(e.Value(), &v) != nil || v.Builder != "next" {
+		t.Fatalf("the key says %+v (%v)", v, err)
+	}
+}
+
+// 1687: the oldest waiting ask is read without being taken: no delivery is counted against it.
+func TestNatsTheOldestWaitingAskIsReadWithoutTakingIt(t *testing.T) {
+	js, _ := aBusWithTheBuckets(t)
+	ctx := t.Context()
+	asks := BuildsOn(js, TheBuildMachine)
+	old := aCheckAskedAgo(45*time.Minute, 1, false)
+	if err := asks.Ask(ctx, old); err != nil {
+		t.Fatal(err)
+	}
+	gate := busGate{conn: js.Conn()}
+	q := &takeQueue{kind: queueOther, worker: broker.CheckWorkers()[1]}
+	at, found, err := gate.OldestWaiting(ctx, q)
+	if err != nil || !found {
+		t.Fatalf("the first waiting ask was not read: %v %v", found, err)
+	}
+	if want, _ := BuildAskedAt(old.ID); !at.Equal(want) {
+		t.Fatalf("read %v, asked %v", at, want)
+	}
+	for range 3 {
+		if _, _, err := gate.OldestWaiting(ctx, q); err != nil {
+			t.Fatal(err)
+		}
+	}
+	info, err := js.Context().ConsumerInfo(q.worker.Stream, q.worker.Name)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if info.Delivered.Consumer != 0 || info.NumAckPending != 0 || info.NumPending != 1 {
+		t.Fatalf("reading the age took the ask: delivered %d, pending acks %d, waiting %d", info.Delivered.Consumer, info.NumAckPending, info.NumPending)
+	}
+	// And an empty queue reads as none.
+	if _, found, err := gate.OldestWaiting(ctx, &takeQueue{kind: queueReady, worker: broker.CheckWorkers()[0]}); err != nil || found {
+		t.Fatalf("an empty queue read as %v (%v)", found, err)
+	}
+}
+
+// Issue 1747: a message that reaches a builder's subscription outside a fetch is not held while the builder runs another ask. A
+// stray check lands in the other queue's subscription of builder A, which then runs a long check; the server hands the stray
+// out again when its wait for an acknowledgement ends (two seconds here), builder B runs it, and A, finished, does not run it a
+// second time from the subscription it held.
+func TestNatsAMessageDeliveredOutsideAFetchIsNotHeldWhileTheBuilderRunsAnotherAsk(t *testing.T) { // 1747
+	js := aBusWithTheBuildRole(t)
+	ctx, stop := context.WithCancel(t.Context())
+	defer stop()
+	for _, w := range broker.CheckWorkers() {
+		w.AckWaitSeconds = 2
+		if err := js.EnsureConsumer(w); err != nil {
+			t.Fatal(err)
+		}
+	}
+	// The long check says it is alive well within the ack wait of two seconds.
+	was := heartbeatWhileWorking
+	heartbeatWhileWorking = 400 * time.Millisecond
+	t.Cleanup(func() { heartbeatWhileWorking = was })
+	asks := BuildsOn(js, TheBuildMachine)
+	long := aPullCheck("build-"+fmt.Sprint(time.Now().UnixNano()), 1, "head-1-0000000", "trunk-0-000000", true)
+	stray := aPullCheck("build-"+fmt.Sprint(time.Now().UnixNano()+1), 2, "head-2-0000000", "trunk-0-000000", false)
+	for _, r := range []BuildRequest{long, stray} {
+		if err := asks.Ask(ctx, r); err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	var mu sync.Mutex
+	runs := map[string][]string{} // ask id -> the machines that ran it
+	release := make(chan struct{})
+	started := make(chan string, 8)
+	do := func(on string) func(context.Context, Build) {
+		return func(ctx context.Context, work Build) {
+			id := work.Request().ID
+			mu.Lock()
+			runs[id] = append(runs[id], on)
+			mu.Unlock()
+			started <- on + ":" + id
+			if id == long.ID {
+				<-release
+			}
+			_ = work.Announce(ctx, BuildResult{ID: id, On: on, Failed: "no"})
+			_ = work.Done()
+		}
+	}
+
+	var a *natsMachine
+	strayLanded := make(chan error, 1)
+	a = MachineOverNATSWith(js, "ace", TheBuildMachine, MachineOptions{Claim: func() bool {
+		// The moment the long check is claimed: the server hands the other queue's waiting check to A's subscription there, as
+		// one that arrives after a fetch gave up does. Done once.
+		select {
+		case <-strayLanded:
+			return true
+		default:
+		}
+		other := a.subs[2]
+		inbox := strings.TrimSuffix(other.sub.Subject, "*") + "stray"
+		body, _ := json.Marshal(map[string]any{"batch": 1, "expires": int64(30 * time.Second)})
+		if err := js.Conn().PublishRequest("$JS.API.CONSUMER.MSG.NEXT."+other.worker.Stream+"."+other.worker.Name, inbox, body); err != nil {
+			strayLanded <- err
+			return true
+		}
+		_ = js.Conn().Flush()
+		for deadline := time.Now().Add(5 * time.Second); time.Now().Before(deadline); time.Sleep(10 * time.Millisecond) {
+			if n, _, _ := other.sub.Pending(); n > 0 {
+				break
+			}
+		}
+		strayLanded <- nil
+		return true
+	}}).(*natsMachine)
+	defer a.Close()
+	go func() { _ = a.Take(ctx, do("ace")) }()
+
+	select {
+	case got := <-started:
+		if got != "ace:"+long.ID {
+			t.Fatalf("started %s, want the long check on ace first", got)
+		}
+	case <-time.After(15 * time.Second):
+		t.Fatal("the long check did not start")
+	}
+	if err := <-strayLanded; err != nil {
+		t.Fatal(err)
+	}
+
+	// B, free, runs the stray once the server hands it out again.
+	b := MachineOverNATS(js, "shanks")
+	defer b.Close()
+	go func() { _ = b.Take(ctx, do("shanks")) }()
+	select {
+	case got := <-started:
+		if got != "shanks:"+stray.ID {
+			t.Fatalf("started %s, want the stray on shanks", got)
+		}
+	case <-time.After(20 * time.Second):
+		t.Fatal("the stray was held by the builder that ran the long check and nobody ran it")
+	}
+	// A finishes, and does not run the stray from the subscription it held.
+	release <- struct{}{}
+	select {
+	case got := <-started:
+		t.Fatalf("%s started after the long check: the stray ran twice", got)
+	case <-time.After(6 * time.Second):
+	}
+	mu.Lock()
+	defer mu.Unlock()
+	if len(runs[stray.ID]) != 1 {
+		t.Fatalf("the stray ran on %v, want once", runs[stray.ID])
+	}
+}
+
+// Review of 1688: a young check pending redelivery (handed back by a Nak) stands ahead of a check older than 30 minutes in the
+// same queue while a plan builds. The let-through fetch would return the young one first, so nothing is fetched: the young
+// check's delivery count stays 1, it is not handed back every second until it is dead-lettered, and it is not run.
+func TestNatsAYoungRedeliveryAheadOfAnOldCheckIsNeverFetchedByTheLetThrough(t *testing.T) {
+	js, plans := aBusWithTheBuckets(t)
+	ctx, stop := context.WithCancel(t.Context())
+	defer stop()
+	asks := BuildsOn(js, TheBuildMachine)
+	young := aCheckAskedAgo(time.Minute, 1, false)
+	if err := asks.Ask(ctx, young); err != nil {
+		t.Fatal(err)
+	}
+	// Someone took it and handed it back: it is pending redelivery.
+	worker := broker.CheckWorkers()[1]
+	sub, err := js.Context().PullSubscribe(worker.Filters[0], worker.Name, nats.Bind(worker.Stream, worker.Name), nats.ManualAck())
+	if err != nil {
+		t.Fatal(err)
+	}
+	got, err := sub.Fetch(1, nats.MaxWait(5*time.Second))
+	if err != nil || len(got) != 1 {
+		t.Fatalf("the young check was not fetched to be handed back: %v", err)
+	}
+	if err := got[0].Nak(); err != nil {
+		t.Fatal(err)
+	}
+	_ = sub.Unsubscribe()
+	old := aCheckAskedAgo(45*time.Minute, 2, false)
+	if err := asks.Ask(ctx, old); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := plans.Put(ctx, "plan-1", []byte(`{"plan":"plan-1"}`)); err != nil {
+		t.Fatal(err)
+	}
+	started := make(chan string, 4)
+	machine := MachineOverNATS(js, "ace")
+	defer machine.Close()
+	go func() {
+		_ = machine.Take(ctx, func(ctx context.Context, work Build) {
+			started <- work.Request().ID
+			_ = work.Announce(ctx, BuildResult{ID: work.Request().ID, On: "ace", Failed: "no"})
+			_ = work.Done()
+		})
+	}()
+	select {
+	case id := <-started:
+		t.Fatalf("%s was run while the young redelivery stood ahead of it", id)
+	case <-time.After(6 * time.Second):
+	}
+	info, err := js.Context().ConsumerInfo(worker.Stream, worker.Name)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if info.Delivered.Consumer != 1 || info.NumRedelivered != 0 {
+		t.Fatalf("the young check was handed out %d time(s), %d redelivered: the let-through fetched it", info.Delivered.Consumer, info.NumRedelivered)
+	}
+}

internal/link/holds.go

@@ -0,0 +1,384 @@
+package link
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/nats-io/nats.go"
+	"github.com/nats-io/nats.go/jetstream"
+
+	"github.com/novox/mesh-controller/internal/broker"
+)
+
+// Checks wait while a delivery builds (novox/hq ADR 0483, issues 1652, 1687 and 1688).
+//
+// **The layout other parts read** (ADR 0484 clause 12 has the checker read the same key; a layout shared between modules
+// belongs in mesh-sdk, and this one is not there yet):
+//
+//   - bucket `mesh-controller_plans-building` (broker.PlansBuildingBucket): one key per delivery plan in its build phase, named
+//     by the plan's id (`plan-<digits>`), value a PlanBuilding as JSON. Written by the controller (lease holder) alone: put when
+//     the plan's first build is asked, put again about every thirty seconds, deleted when the phase ends; the bucket's age
+//     (two minutes) removes the key of a controller that stopped. A reader lists the keys and needs no value to decide.
+//   - key `let-through` in bucket `mesh-controller_checks-running` (broker.CheckRunningBucket), value a LetThrough as JSON:
+//     created by the holder that takes a check which waited LetThroughAfter, refreshed at each heartbeat of that check, deleted
+//     when it ends. At most one such check runs at a time on the build seat, because the create fails while the key is there.
+
+// LetThroughAfter is how long a check waits, while a delivery builds, before it is let through (ADR 0483 clause 3).
+const LetThroughAfter = 30 * time.Minute
+
+// LetThroughKey is the key of the one check let through, in the bucket of the checks running. It has no dot, so it is no
+// pull request's key (`<owner>.<repo>.<number>`) and no group's (`group.<name>`).
+const LetThroughKey = "let-through"
+
+// PlanBuilding is the value of a plan's key in the bucket of the plans building.
+type PlanBuilding struct {
+	Plan  string    `json:"plan"`
+	Since time.Time `json:"since"`
+	Seen  time.Time `json:"seen"`
+}
+
+// LetThrough is the value of the let-through key: the builder that runs the check and when it took it.
+type LetThrough struct {
+	Builder string    `json:"builder"`
+	Since   time.Time `json:"since"`
+}
+
+// checkGate is what a taking loop asks of the bus to hold checks while a delivery builds; a variable of the machine so a test
+// hands out against fakes.
+type checkGate interface {
+	// Building is the ids of the plans in their build phase: the keys of the bucket. An error is a bucket that cannot be read.
+	Building(ctx context.Context) ([]string, error)
+	// OldestWaiting is when the first waiting ask of a check queue was asked, read without taking it.
+	OldestWaiting(ctx context.Context, q *takeQueue) (time.Time, bool, error)
+	// TakeLetThrough creates the let-through key for a builder: false when it is there (another builder runs a check let through).
+	TakeLetThrough(ctx context.Context, builder string) (bool, error)
+	// KeepLetThrough puts the key again, and DropLetThrough deletes it.
+	KeepLetThrough(ctx context.Context, builder string) error
+	DropLetThrough(ctx context.Context) error
+}
+
+// busGate is the gate over a connection.
+type busGate struct{ conn *nats.Conn }
+
+func (g busGate) Building(ctx context.Context) ([]string, error) {
+	return BuildingPlans(ctx, g.conn)
+}
+
+// BuildingPlans is the ids of the delivery plans in their build phase, from the controller's bucket. A reader that is not the
+// build seat's (the checker, ADR 0484 clause 12) calls this. A bucket that cannot be read is an error; none is none building.
+func BuildingPlans(ctx context.Context, conn *nats.Conn) ([]string, error) {
+	api, err := jetstream.New(conn)
+	if err != nil {
+		return nil, err
+	}
+	kv, err := api.KeyValue(ctx, broker.PlansBuildingBucket)
+	if err != nil {
+		return nil, fmt.Errorf("the plans building: %w", err)
+	}
+	keys, err := kv.Keys(ctx)
+	if errors.Is(err, jetstream.ErrNoKeysFound) {
+		return nil, nil
+	}
+	if err != nil {
+		return nil, fmt.Errorf("the plans building: %w", err)
+	}
+	return keys, nil
+}
+
+// OldestWaiting is when the check a fetch of this queue would return was asked, read without taking anything.
+//
+// **It sees redeliveries** (review of issue 1688). A work queue keeps only what is not acknowledged: what the worker handed out
+// (sequence up to its delivered mark) is running or waiting to be handed out again, and a fetch returns those before anything
+// new. An ask there that is not said running (a key of the checks running names its build) is such a redelivery. If one is
+// younger than the bound, the queue's next fetch would return it, and a young check is never let through: the queue is not a
+// candidate, and nothing is fetched, so nothing is handed back and no delivery is spent. Otherwise the candidate is the oldest
+// of the redeliveries and the first ask after the delivered mark.
+func (g busGate) OldestWaiting(ctx context.Context, q *takeQueue) (time.Time, bool, error) {
+	info, err := jsOf(g.conn).ConsumerInfo(q.worker.Stream, q.worker.Name, nats.Context(ctx))
+	if err != nil {
+		return time.Time{}, false, fmt.Errorf("the %s worker: %w", q.kind, err)
+	}
+	var oldest time.Time
+	note := func(at time.Time) {
+		if oldest.IsZero() || at.Before(oldest) {
+			oldest = at
+		}
+	}
+	for seq := uint64(1); seq <= info.Delivered.Stream; {
+		data, stored, got, err := g.directGet(ctx, q, seq)
+		if err != nil {
+			return time.Time{}, false, err
+		}
+		if got == 0 || got > info.Delivered.Stream {
+			break
+		}
+		seq = got + 1
+		var r BuildRequest
+		if json.Unmarshal(data, &r) != nil {
+			continue
+		}
+		running, ok, err := ReadRunning(ctx, g.conn, r)
+		if err != nil {
+			return time.Time{}, false, fmt.Errorf("the checks running: %w", err)
+		}
+		if ok && running.Build == r.ID {
+			continue // running now
+		}
+		at := askedAtOf(data, stored)
+		if time.Since(at) < LetThroughAfter {
+			return time.Time{}, false, nil // a young redelivery is next in line
+		}
+		note(at)
+	}
+	data, stored, got, err := g.directGet(ctx, q, info.Delivered.Stream+1)
+	if err != nil {
+		return time.Time{}, false, err
+	}
+	if got != 0 {
+		note(askedAtOf(data, stored))
+	}
+	return oldest, !oldest.IsZero(), nil
+}
+
+// directGet reads the first ask at or after a sequence of a check queue's stream, without taking it: its body, the time the
+// stream stored it, and its sequence (0 when there is none).
+func (g busGate) directGet(ctx context.Context, q *takeQueue, seq uint64) ([]byte, string, uint64, error) {
+	body, _ := json.Marshal(map[string]any{"seq": seq, "next_by_subj": q.worker.Filters[0]})
+	reply, err := g.conn.RequestWithContext(ctx, "$JS.API.DIRECT.GET."+q.worker.Stream, body)
+	if err != nil {
+		return nil, "", 0, fmt.Errorf("the asks of the %s queue: %w", q.kind, err)
+	}
+	if st := reply.Header.Get("Status"); st == "404" || st == "408" {
+		return nil, "", 0, nil
+	}
+	got, _ := strconv.ParseUint(reply.Header.Get("Nats-Sequence"), 10, 64)
+	return reply.Data, reply.Header.Get("Nats-Time-Stamp"), got, nil
+}
+
+// askedAtOf is when a request was asked: its id says (NewBuildID), the stream's own time otherwise, and a request that says
+// neither is taken as asked now, so it is never let through on a guess.
+func askedAtOf(data []byte, stored string) time.Time {
+	var r BuildRequest
+	if json.Unmarshal(data, &r) == nil {
+		if at, ok := BuildAskedAt(r.ID); ok {
+			return at
+		}
+	}
+	if at, err := time.Parse(time.RFC3339Nano, stored); err == nil {
+		return at
+	}
+	return time.Now()
+}
+
+func jsOf(conn *nats.Conn) nats.JetStreamContext {
+	js, _ := conn.JetStream()
+	return js
+}
+
+func (g busGate) TakeLetThrough(ctx context.Context, builder string) (bool, error) {
+	kv, err := runningBucket(ctx, g.conn)
+	if err != nil {
+		return false, err
+	}
+	body, _ := json.Marshal(LetThrough{Builder: builder, Since: time.Now().UTC()})
+	if _, err := kv.Create(ctx, LetThroughKey, body); err != nil {
+		if errors.Is(err, jetstream.ErrKeyExists) {
+			return false, nil
+		}
+		return false, err
+	}
+	return true, nil
+}
+
+func (g busGate) KeepLetThrough(ctx context.Context, builder string) error {
+	kv, err := runningBucket(ctx, g.conn)
+	if err != nil {
+		return err
+	}
+	body, _ := json.Marshal(LetThrough{Builder: builder, Since: time.Now().UTC()})
+	_, err = kv.Put(ctx, LetThroughKey, body)
+	return err
+}
+
+func (g busGate) DropLetThrough(ctx context.Context) error {
+	kv, err := runningBucket(ctx, g.conn)
+	if err != nil {
+		return err
+	}
+	if err := kv.Delete(ctx, LetThroughKey); err != nil && !errors.Is(err, jetstream.ErrKeyNotFound) {
+		return err
+	}
+	return nil
+}
+
+// CheckHold is what a reader of the queue (the controller's own verb) says of why checks wait: the plans building and the check
+// let through now, if one.
+type CheckHold struct {
+	Plans      []string
+	LetThrough *LetThrough
+}
+
+// ReadCheckHold reads the plans building and the let-through key. A bucket that cannot be read is an error the caller says.
+func ReadCheckHold(ctx context.Context, conn *nats.Conn) (CheckHold, error) {
+	var h CheckHold
+	plans, err := BuildingPlans(ctx, conn)
+	if err != nil {
+		return h, err
+	}
+	h.Plans = plans
+	kv, err := runningBucket(ctx, conn)
+	if err != nil {
+		return h, err
+	}
+	e, err := kv.Get(ctx, LetThroughKey)
+	if errors.Is(err, jetstream.ErrKeyNotFound) {
+		return h, nil
+	}
+	if err != nil {
+		return h, err
+	}
+	var l LetThrough
+	if json.Unmarshal(e.Value(), &l) == nil {
+		h.LetThrough = &l
+	}
+	return h, nil
+}
+
+// gateOf is the machine's gate: the one a test set, or the bus's.
+func (m *natsMachine) gateOf() checkGate {
+	if m.gate != nil {
+		return m.gate
+	}
+	return busGate{conn: m.js.Conn()}
+}
+
+// sayOnce says a fault of the hold the first time it happens and again once it has recovered and happened anew, so a bucket
+// that cannot be read is in the log once and not every second.
+func (m *natsMachine) sayOnce(kind string, err error) {
+	if m.said == nil {
+		m.said = map[string]bool{}
+	}
+	if err == nil {
+		delete(m.said, kind)
+		return
+	}
+	if !m.said[kind] {
+		m.said[kind] = true
+		fmt.Fprintf(os.Stderr, "%s; read as empty, so no check is held for it: %v\n", kind, err)
+	}
+}
+
+// building says whether a delivery builds: any key in the bucket of the plans building. **A bucket that cannot be read is said
+// and read as empty** (ADR 0483 clause 2): a check wrongly held costs a pull request its turn, a fault that held every check
+// would stop the mesh checking.
+func (m *natsMachine) building(ctx context.Context) []string {
+	plans, err := m.gateOf().Building(ctx)
+	m.sayOnce("the bucket of the plans building cannot be read", err)
+	if err != nil {
+		// empty-on-error: ADR 0483 clause 2 — a bucket that cannot be read is said (sayOnce) and read as empty, because a check wrongly held costs a pull request its turn and a fault that held every check would stop the mesh checking.
+		return nil
+	}
+	return plans
+}
+
+// letOneThrough takes the oldest waiting check of the two check queues once it waited LetThroughAfter, by the one let-through
+// key: nil when none waited that long, when another holder runs the check let through, or when the check was taken by
+// another holder in the moment. The caller owns the key while it returns a message and releases it with dropLetThrough.
+func (m *natsMachine) letOneThrough(ctx context.Context) (*nats.Msg, *takeQueue) {
+	gate := m.gateOf()
+	var oldest *takeQueue
+	var oldestAt time.Time
+	for _, q := range m.subs[1:] {
+		if !q.bound {
+			continue
+		}
+		at, found, err := gate.OldestWaiting(ctx, q)
+		m.sayOnce("the first waiting check of the "+q.kind+" queue cannot be read", err)
+		if err != nil || !found {
+			continue
+		}
+		if oldest == nil || at.Before(oldestAt) {
+			oldest, oldestAt = q, at
+		}
+	}
+	if oldest == nil || time.Since(oldestAt) < m.letThroughAfter() {
+		return nil, nil
+	}
+	created, err := gate.TakeLetThrough(ctx, m.on)
+	m.sayOnce("the let-through key cannot be taken", err)
+	if err != nil || !created {
+		return nil, nil
+	}
+	trying, end := context.WithTimeout(ctx, checkTakeWait)
+	defer end()
+	fetched, err := oldest.pull.Fetch(1, nats.Context(trying))
+	if len(fetched) == 0 {
+		if err != nil && !errors.Is(err, context.DeadlineExceeded) && !errors.Is(err, nats.ErrTimeout) && !errors.Is(err, context.Canceled) {
+			m.sayOnce("the "+oldest.kind+" check queue cannot be taken from", err)
+		}
+		m.dropLetThrough(ctx)
+		return nil, nil
+	}
+	// Another holder took the one that was oldest in the moment between the read and the fetch: what came is a younger check,
+	// handed back and the key released, never run during a delivery.
+	if askedAtOf(fetched[0].Data, "").Add(m.letThroughAfter()).After(time.Now()) {
+		_ = fetched[0].NakWithDelay(time.Minute) // a delay keeps it off the queue of redeliveries a fetch reads first
+		m.dropLetThrough(ctx)
+		return nil, nil
+	}
+	m.letKey = true
+	return fetched[0], oldest
+}
+
+func (m *natsMachine) letThroughAfter() time.Duration {
+	if m.opts.LetThroughAfter > 0 {
+		return m.opts.LetThroughAfter
+	}
+	return LetThroughAfter
+}
+
+// dropLetThrough deletes the let-through key this holder created.
+func (m *natsMachine) dropLetThrough(ctx context.Context) {
+	m.letKey = false
+	if err := m.gateOf().DropLetThrough(context.WithoutCancel(ctx)); err != nil {
+		fmt.Fprintf(os.Stderr, "could not take the let-through key away; it ages out: %v\n", err)
+	}
+}
+
+// keepLetThrough refreshes the key at every heartbeat while the check it was created for runs, and returns the end of that: it
+// stops the refreshing and deletes the key.
+func (m *natsMachine) keepLetThrough(ctx context.Context, every time.Duration) (end func()) {
+	done := make(chan struct{})
+	finished := make(chan struct{})
+	go func() {
+		defer close(finished)
+		tick := time.NewTicker(every)
+		defer tick.Stop()
+		for {
+			select {
+			case <-done:
+				return
+			case <-tick.C:
+				putting, stop := context.WithTimeout(context.WithoutCancel(ctx), 5*time.Second)
+				if err := m.gateOf().KeepLetThrough(putting, m.on); err != nil {
+					fmt.Fprintf(os.Stderr, "could not keep the let-through key: %v\n", err)
+				}
+				stop()
+			}
+		}
+	}()
+	return func() {
+		close(done)
+		<-finished
+		m.dropLetThrough(ctx)
+	}
+}
+
+// whyHeld words the plans a check is held for, for the log.
+func whyHeld(plans []string) string { return strings.Join(plans, ", ") }

internal/link/holds_test.go

@@ -0,0 +1,307 @@
+package link
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/nats-io/nats.go"
+
+	"github.com/novox/mesh-controller/internal/broker"
+)
+
+// Checks wait while a delivery builds (novox/hq ADR 0483, issues 1687 and 1688): the order the taking loop hands out in, against
+// fake queues and a fake bus.
+
+// fakeQueue is a queue the loop fetches from: the asks waiting, handed out in order, and how often it was fetched from.
+type fakeQueue struct {
+	mu      sync.Mutex
+	waiting []*nats.Msg
+	fetched int
+}
+
+func (f *fakeQueue) Fetch(int, ...nats.PullOpt) ([]*nats.Msg, error) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	f.fetched++
+	if len(f.waiting) == 0 {
+		return nil, nats.ErrTimeout
+	}
+	m := f.waiting[0]
+	f.waiting = f.waiting[1:]
+	return []*nats.Msg{m}, nil
+}
+
+func (f *fakeQueue) fetches() int {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return f.fetched
+}
+
+// askedAgo is a request asked this long ago, as the id says, for a check of this pull request.
+func askedAgo(ago time.Duration, number int, ready bool) *nats.Msg {
+	id := fmt.Sprintf("build-%d", time.Now().Add(-ago).UnixNano())
+	body, _ := json.Marshal(BuildRequest{ID: id, Repository: "r", Check: &CheckRequest{Owner: "o", Repo: "r", Number: number, Ready: ready}})
+	return &nats.Msg{Data: body}
+}
+
+func buildAsk() *nats.Msg {
+	body, _ := json.Marshal(BuildRequest{ID: fmt.Sprintf("build-%d", time.Now().UnixNano()), Repository: "r"})
+	return &nats.Msg{Data: body}
+}
+
+// fakeGate is the bus as the loop sees it for the hold.
+type fakeGate struct {
+	mu         sync.Mutex
+	plans      []string
+	plansErr   error
+	letKey     bool
+	taken      int
+	kept       int
+	dropped    int
+	oldestFrom map[*takeQueue]*fakeQueue
+}
+
+func (g *fakeGate) Building(context.Context) ([]string, error) {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	return g.plans, g.plansErr
+}
+
+// OldestWaiting reads the first waiting ask of the fake queue without taking it.
+func (g *fakeGate) OldestWaiting(_ context.Context, q *takeQueue) (time.Time, bool, error) {
+	f := g.oldestFrom[q]
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	if len(f.waiting) == 0 {
+		return time.Time{}, false, nil
+	}
+	return askedAtOf(f.waiting[0].Data, ""), true, nil
+}
+
+func (g *fakeGate) TakeLetThrough(context.Context, string) (bool, error) {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	if g.letKey {
+		return false, nil
+	}
+	g.letKey = true
+	g.taken++
+	return true, nil
+}
+
+func (g *fakeGate) KeepLetThrough(context.Context, string) error {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	g.kept++
+	return nil
+}
+
+func (g *fakeGate) DropLetThrough(context.Context) error {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	g.letKey = false
+	g.dropped++
+	return nil
+}
+
+func (g *fakeGate) setPlans(plans ...string) {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	g.plans = plans
+}
+
+// aMachine is a loop on the build seat with a delivery queue and the two check queues, all fake.
+func aMachine() (m *natsMachine, builds, ready, other *fakeQueue, gate *fakeGate) {
+	builds, ready, other = &fakeQueue{}, &fakeQueue{}, &fakeQueue{}
+	subs := []*takeQueue{
+		{kind: queueBuilds, pull: builds, bound: true},
+		{kind: queueReady, pull: ready, bound: true},
+		{kind: queueOther, pull: other, bound: true},
+	}
+	gate = &fakeGate{oldestFrom: map[*takeQueue]*fakeQueue{subs[1]: ready, subs[2]: other}}
+	m = &natsMachine{on: "ace", seat: broker.TheBuildSeat, subs: subs, gate: gate}
+	return m, builds, ready, other, gate
+}
+
+func nextAsk(t *testing.T, m *natsMachine) (*nats.Msg, *takeQueue) {
+	t.Helper()
+	msg, q, err := m.next(context.Background())
+	if err != nil {
+		t.Fatal(err)
+	}
+	return msg, q
+}
+
+// 1652/AC6, 1687: a waiting check is not taken while a plan key is there, also when the delivery's queue is empty for the moment
+// (between two tiers).
+func TestNoCheckIsTakenWhileAPlanBuilds(t *testing.T) {
+	m, builds, ready, other, gate := aMachine()
+	gate.setPlans("plan-1")
+	ready.waiting = []*nats.Msg{askedAgo(2*time.Minute, 1, true)}
+	other.waiting = []*nats.Msg{askedAgo(5*time.Minute, 2, false)}
+	// The delivery's queue is empty: between two tiers.
+	if msg, _ := nextAsk(t, m); msg != nil {
+		t.Fatalf("a check was taken while plan-1 builds: %s", msg.Data)
+	}
+	if ready.fetches() != 0 || other.fetches() != 0 {
+		t.Fatalf("the check queues were fetched from while plan-1 builds: ready %d, other %d", ready.fetches(), other.fetches())
+	}
+	if builds.fetches() == 0 {
+		t.Fatal("the delivery's queue was not looked at")
+	}
+}
+
+// A delivery build is still taken first, and the check waits behind it.
+func TestADeliveryBuildIsTakenWhileAPlanBuilds(t *testing.T) {
+	m, builds, ready, _, gate := aMachine()
+	gate.setPlans("plan-1")
+	builds.waiting = []*nats.Msg{buildAsk()}
+	ready.waiting = []*nats.Msg{askedAgo(2*time.Minute, 1, true)}
+	msg, q := nextAsk(t, m)
+	if msg == nil || q.kind != queueBuilds {
+		t.Fatalf("the delivery build was not taken first: %v %v", msg, q)
+	}
+	if ready.fetches() != 0 {
+		t.Fatal("a check queue was fetched from")
+	}
+}
+
+// It is taken once the key is gone (deleted at the phase's end, or expired).
+func TestACheckIsTakenOnceTheKeyIsGone(t *testing.T) {
+	m, _, ready, _, gate := aMachine()
+	gate.setPlans("plan-1")
+	ready.waiting = []*nats.Msg{askedAgo(2*time.Minute, 1, true)}
+	if msg, _ := nextAsk(t, m); msg != nil {
+		t.Fatal("taken while the key is there")
+	}
+	gate.setPlans()
+	msg, q := nextAsk(t, m)
+	if msg == nil || q.kind != queueReady {
+		t.Fatalf("the check was not taken once the key was gone: %v", q)
+	}
+	if m.letKey {
+		t.Fatal("a check taken without a delivery building took the let-through key")
+	}
+}
+
+// 1687: a bucket that cannot be read is said and read as empty, so the checks are taken.
+func TestAnUnreadableBucketHoldsNoCheck(t *testing.T) {
+	m, _, ready, _, gate := aMachine()
+	gate.plansErr = errors.New("permissions violation")
+	gate.plans = []string{"plan-1"}
+	ready.waiting = []*nats.Msg{askedAgo(2*time.Minute, 1, true)}
+	msg, q := nextAsk(t, m)
+	if msg == nil || q.kind != queueReady {
+		t.Fatal("a check was held because the bucket could not be read")
+	}
+	if !m.said["the bucket of the plans building cannot be read"] {
+		t.Fatalf("the fault was not said: %v", m.said)
+	}
+}
+
+// 1688: a check asked 30 minutes ago is let through while a key is there; the oldest across both queues by ask time, ahead of a
+// waiting delivery build; a younger one is never taken before it.
+func TestTheOldestCheckPastTheBoundGoesFirstAcrossBothQueues(t *testing.T) {
+	m, builds, ready, other, gate := aMachine()
+	gate.setPlans("plan-1")
+	builds.waiting = []*nats.Msg{buildAsk()}
+	ready.waiting = []*nats.Msg{askedAgo(31*time.Minute, 1, true)}
+	other.waiting = []*nats.Msg{askedAgo(50*time.Minute, 2, false)}
+	msg, q := nextAsk(t, m)
+	if msg == nil || q.kind != queueOther || !strings.Contains(string(msg.Data), `"number":2`) {
+		t.Fatalf("the oldest check (50 minutes, other queue) was not let through first: %v", q)
+	}
+	if !m.letKey || gate.taken != 1 {
+		t.Fatalf("the let-through key was not created for it: held %v, created %d", m.letKey, gate.taken)
+	}
+}
+
+// 1688: several checks past the bound go one at a time, oldest first: while the key is there every other holder takes builds.
+func TestChecksPastTheBoundGoOneAtATimeOldestFirst(t *testing.T) {
+	m, builds, ready, other, gate := aMachine()
+	gate.setPlans("plan-1")
+	ready.waiting = []*nats.Msg{askedAgo(33*time.Minute, 1, true), askedAgo(31*time.Minute, 3, true)}
+	other.waiting = []*nats.Msg{askedAgo(40*time.Minute, 2, false)}
+	var order []string
+	for range 3 {
+		msg, q := nextAsk(t, m)
+		if msg == nil {
+			t.Fatalf("no check let through after %v", order)
+		}
+		var r BuildRequest
+		_ = json.Unmarshal(msg.Data, &r)
+		order = append(order, fmt.Sprint(r.Check.Number, q.kind))
+		// While it runs another holder (the same fake gate, key held) takes builds only.
+		builds.waiting = []*nats.Msg{buildAsk()}
+		other2 := &natsMachine{on: "shanks", seat: broker.TheBuildSeat, subs: m.subs, gate: gate}
+		if next, nq := nextAsk(t, other2); next == nil || nq.kind != queueBuilds {
+			t.Fatalf("a second holder did not take a delivery build while the let-through key is held: %v", nq)
+		}
+		m.dropLetThrough(context.Background())
+	}
+	if want := []string{"2other", "1ready", "3ready"}; strings.Join(order, " ") != strings.Join(want, " ") {
+		t.Fatalf("checks went %v, want %v (oldest first, one at a time)", order, want)
+	}
+}
+
+// 1688: a check younger than the bound is never let through, and no key is created for it.
+func TestACheckYoungerThanTheBoundIsHeld(t *testing.T) {
+	m, _, ready, _, gate := aMachine()
+	gate.setPlans("plan-1")
+	ready.waiting = []*nats.Msg{askedAgo(29*time.Minute, 1, true)}
+	if msg, _ := nextAsk(t, m); msg != nil {
+		t.Fatal("a check of 29 minutes was let through")
+	}
+	if gate.taken != 0 {
+		t.Fatal("a let-through key was created for a young check")
+	}
+	if ready.fetches() != 0 {
+		t.Fatal("the age was read by a fetch; it must be read without taking")
+	}
+}
+
+// 1688: the holder that lost the race for the key, or the check, takes builds only; and a younger check that arrives in the
+// moment between the read and the fetch is handed back with the key released.
+func TestAYoungerCheckFetchedInTheRaceIsHandedBackAndTheKeyReleased(t *testing.T) {
+	m, _, ready, _, gate := aMachine()
+	gate.setPlans("plan-1")
+	// The read sees the old check; before the fetch another holder takes it and a young one is first.
+	gate.oldestFrom[m.subs[1]] = &fakeQueue{waiting: []*nats.Msg{askedAgo(40*time.Minute, 1, true)}}
+	ready.waiting = []*nats.Msg{askedAgo(time.Minute, 2, true)}
+	if msg, _ := nextAsk(t, m); msg != nil {
+		t.Fatalf("a young check was run during a delivery: %s", msg.Data)
+	}
+	if gate.letKey || m.letKey {
+		t.Fatal("the let-through key was left held")
+	}
+}
+
+// The key of a check let through is refreshed while it runs and deleted when it ends.
+func TestTheLetThroughKeyLivesAsLongAsTheCheckRuns(t *testing.T) {
+	m, _, ready, _, gate := aMachine()
+	gate.setPlans("plan-1")
+	ready.waiting = []*nats.Msg{askedAgo(40*time.Minute, 1, true)}
+	if msg, _ := nextAsk(t, m); msg == nil {
+		t.Fatal("not let through")
+	}
+	end := m.keepLetThrough(context.Background(), 20*time.Millisecond)
+	time.Sleep(120 * time.Millisecond)
+	gate.mu.Lock()
+	held, kept := gate.letKey, gate.kept
+	gate.mu.Unlock()
+	if !held || kept == 0 {
+		t.Errorf("while the check ran the key was held %v and refreshed %d times", held, kept)
+	}
+	end()
+	gate.mu.Lock()
+	defer gate.mu.Unlock()
+	if gate.letKey || m.letKey {
+		t.Error("the key was not deleted when the check ended")
+	}
+}

internal/link/queue.go

@@ -88,6 +88,40 @@ type Queue struct {
 	AckPending int         `json:"ack-pending"`
 	MaxDeliver int         `json:"max-deliver"`
 	Asks       []QueuedAsk `json:"asks"`
+	// Hold is why the checks wait: the delivery plans building and the check let through now (ADR 0483, issues 1687 and 1688);
+	// HoldFault says the bucket could not be read, in which case Hold is empty.
+	Hold      CheckHold `json:"hold"`
+	HoldFault string    `json:"hold-fault,omitempty"`
+}
+
+// LetThroughOf says which ask is let through under ADR 0483 clause 3: the check that waited LetThroughAfter or longer and is next
+// (the oldest waiting check across the two check queues, while a delivery builds), and the check in flight that was let through
+// (the let-through key names the machine that runs it). Empty when none.
+func (q Queue) LetThroughOf(now time.Time) (next, running string) {
+	if len(q.Hold.Plans) == 0 && q.Hold.LetThrough == nil {
+		return "", ""
+	}
+	var oldest *QueuedAsk
+	for i := range q.Asks {
+		a := &q.Asks[i]
+		if a.State != AskWaiting || a.Request.Check == nil || a.AskedAt.IsZero() {
+			continue
+		}
+		if oldest == nil || a.AskedAt.Before(oldest.AskedAt) {
+			oldest = a
+		}
+	}
+	if oldest != nil && len(q.Hold.Plans) > 0 && now.Sub(oldest.AskedAt) >= LetThroughAfter {
+		next = oldest.ID
+	}
+	if q.Hold.LetThrough != nil {
+		for _, a := range q.Asks {
+			if a.State == AskInFlight && a.Request.Check != nil && a.On == q.Hold.LetThrough.Builder {
+				running = a.ID
+			}
+		}
+	}
+	return next, running
 }
 
 // Of is the asks in one state, in queue order.
@@ -319,6 +353,16 @@ func ReadQueue(ctx context.Context, js *broker.JetStream, seat string) (Queue, e
 	for _, r := range reads {
 		q.Asks = append(q.Asks, ClassifyQueue(r.asks, r.delivered, r.ackPending, heard)...)
 	}
+	// And why the checks wait, for the build seat that has them (ADR 0483 clause 5). A bucket that cannot be read is said by the
+	// queue, which then names no plan.
+	if seat == broker.TheBuildSeat {
+		hold, err := ReadCheckHold(ctx, js.Conn())
+		if err != nil {
+			q.HoldFault = err.Error()
+		} else {
+			q.Hold = hold
+		}
+	}
 	return q, nil
 }
 

internal/link/stray_test.go

@@ -0,0 +1,54 @@
+package link
+
+import (
+	"testing"
+	"time"
+
+	"github.com/nats-io/nats.go"
+)
+
+// Issue 1747: a builder never holds a message it does not start. A message the server sends after a fetch gave up lands in that
+// queue's subscription; while an ask runs, the queues it does not come from are let go of, so such a message is not held
+// unheartbeated for the whole run, and the queues are bound again before the next fetch.
+
+func TestARunningAskLetsGoOfTheQueuesItDidNotComeFrom(t *testing.T) {
+	m, builds, _, other, _ := aMachine()
+	// A delivery build reached the builds queue's subscription after its fetch gave up: it sits in the buffer.
+	builds.waiting = []*nats.Msg{buildAsk()}
+	took := m.subs[1] // the running ask came from the ready queue
+
+	m.quiesce(took)
+
+	if !took.bound || took.pull == nil {
+		t.Error("the queue the running ask came from was let go of; its acknowledgements need it")
+	}
+	for _, q := range []*takeQueue{m.subs[0], m.subs[2]} {
+		if q.bound || q.pull != nil || q.sub != nil {
+			t.Errorf("the %s queue is still held while an ask runs", q.kind)
+		}
+	}
+	// Nothing is fetched from a queue that was let go of: a message in it cannot be taken unheartbeated.
+	if msg, _, _ := m.next(t.Context()); msg != nil {
+		t.Fatalf("a message was taken from a queue that was let go of: %s", msg.Data)
+	}
+	if builds.fetches() != 0 || other.fetches() != 0 {
+		t.Errorf("fetched from a queue that was let go of: builds %d, other %d", builds.fetches(), other.fetches())
+	}
+}
+
+// Once the ask is settled the queue it came from is let go of too: a second message the server delivered into its subscription
+// on a pull request still open is not run after the ask ends (review of 1747).
+func TestTheQueueTheAskCameFromIsLetGoOfWhenTheAskEnds(t *testing.T) {
+	m, _, ready, _, _ := aMachine()
+	ready.waiting = []*nats.Msg{askedAgo(time.Minute, 2, true)}
+	m.quiesce(nil)
+	for _, q := range m.subs {
+		if q.bound || q.pull != nil {
+			t.Errorf("the %s queue is still held after the ask ended", q.kind)
+		}
+	}
+	m.subs[0].bound, m.subs[0].pull = true, &fakeQueue{}
+	if msg, _, _ := m.next(t.Context()); msg != nil || ready.fetches() != 0 {
+		t.Fatal("a message held in the kept queue's subscription was taken after the ask ended")
+	}
+}