diff --git a/docs/runner-protocol-roadmap.md b/docs/runner-protocol-roadmap.md
index 5b66a88..97a45ab 100644
--- a/docs/runner-protocol-roadmap.md
+++ b/docs/runner-protocol-roadmap.md
@@ -51,6 +51,9 @@ controller 使用单一 Go 二进制；默认在同一进程启用 `scheduler`
   SPIFFE ID写入 sandbox metadata；通过 `extensions.poolRef=ci-vm` 使用既有 Kata Pool。
 - 结果处理顺序固定为回报 Gitea、清理后端、ACK assignment；重投时先查询 Gitea 终态，
   从而关闭后端已删除但消息尚未 ACK 的崩溃窗口。
+- pod 与 vm 使用独立 durable consumer 和并发上限。worker 持有消息期间持续 reconcile
+  后端并发送 `InProgress`；只有完整完成才 `DoubleAck`，进程退出则保留未确认消息供
+  其他实例恢复，临时后端错误使用延迟 NAK。
 - Pod 与 VM 共享 task/executor 协议，只有环境创建和销毁实现不同。
 
 ## 实现顺序
diff --git a/internal/assignmentqueue/jetstream.go b/internal/assignmentqueue/jetstream.go
index 25e0c57..59cc49f 100644
--- a/internal/assignmentqueue/jetstream.go
+++ b/internal/assignmentqueue/jetstream.go
@@ -6,6 +6,7 @@ import (
 	"errors"
 	"fmt"
 	"strings"
+	"sync"
 	"time"
 
 	"github.com/nats-io/nats.go"
@@ -62,9 +63,10 @@ type Message interface {
 
 // Processor maps one delivery to one idempotent worker reconciliation.
 type Processor struct {
-	TrustDomain string
-	Handler     Handler
-	RetryDelay  time.Duration
+	TrustDomain  string
+	Handler      Handler
+	RetryDelay   time.Duration
+	PollInterval time.Duration
 }
 
 func (p Processor) Process(ctx context.Context, message Message) error {
@@ -94,3 +96,125 @@ func (p Processor) Process(ctx context.Context, message Message) error {
 	}
 	return nil
 }
+
+// ProcessUntilDone holds one durable delivery while repeatedly reconciling
+// backend state. Cancellation leaves it unacknowledged for another process.
+func (p Processor) ProcessUntilDone(ctx context.Context, message Message) error {
+	if p.Handler == nil {
+		return errors.New("assignment handler is required")
+	}
+	assignment, err := taskassignment.Unmarshal(message.Data(), p.TrustDomain)
+	if err != nil {
+		return errors.Join(err, message.TermWithReason("invalid assignment"))
+	}
+	interval := p.PollInterval
+	if interval <= 0 {
+		interval = 2 * time.Second
+	}
+	for {
+		done, handleErr := p.Handler.Handle(ctx, assignment)
+		if handleErr != nil {
+			delay := p.RetryDelay
+			if delay <= 0 {
+				delay = 15 * time.Second
+			}
+			return errors.Join(handleErr, message.NakWithDelay(delay))
+		}
+		if done {
+			if err := message.DoubleAck(ctx); err != nil {
+				return fmt.Errorf("ack assignment %s: %w", assignment.ID, err)
+			}
+			return nil
+		}
+		if err := message.InProgress(); err != nil {
+			return fmt.Errorf("extend assignment %s acknowledgement: %w", assignment.ID, err)
+		}
+		timer := time.NewTimer(interval)
+		select {
+		case <-ctx.Done():
+			if !timer.Stop() {
+				<-timer.C
+			}
+			return ctx.Err()
+		case <-timer.C:
+		}
+	}
+}
+
+type consumeAPI interface {
+	Consume(jetstream.MessageHandler, ...jetstream.PullConsumeOpt) (jetstream.ConsumeContext, error)
+}
+
+// ConsumerComponent runs bounded reconciliation goroutines for one durable
+// backend consumer. The goroutine set is operational state, not task storage.
+type ConsumerComponent struct {
+	Consumer  consumeAPI
+	Processor Processor
+	Capacity  int
+	OnError   func(error)
+}
+
+type consumerManager interface {
+	CreateOrUpdateConsumer(context.Context, string, jetstream.ConsumerConfig) (jetstream.Consumer, error)
+}
+
+// OpenConsumer creates the durable backend cursor. Capacity is enforced both
+// server-side and by ConsumerComponent's local semaphore.
+func OpenConsumer(ctx context.Context, manager consumerManager, stream, subjectBase string, backend taskassignment.Backend, capacity int) (jetstream.Consumer, error) {
+	if manager == nil || stream == "" || strings.TrimSuffix(subjectBase, ".") == "" || capacity < 1 {
+		return nil, errors.New("JetStream manager, stream, subject base, and positive capacity are required")
+	}
+	if backend != taskassignment.BackendPod && backend != taskassignment.BackendVM {
+		return nil, fmt.Errorf("unsupported assignment backend %q", backend)
+	}
+	consumer, err := manager.CreateOrUpdateConsumer(ctx, stream, jetstream.ConsumerConfig{
+		Name:          string(backend),
+		Durable:       string(backend),
+		FilterSubject: strings.TrimSuffix(subjectBase, ".") + "." + string(backend),
+		AckPolicy:     jetstream.AckExplicitPolicy,
+		AckWait:       5 * time.Minute,
+		MaxAckPending: capacity,
+		MaxDeliver:    20,
+	})
+	if err != nil {
+		return nil, fmt.Errorf("open %s assignment consumer: %w", backend, err)
+	}
+	return consumer, nil
+}
+
+func (c ConsumerComponent) Run(ctx context.Context) error {
+	if c.Consumer == nil || c.Capacity < 1 {
+		return errors.New("JetStream consumer and positive capacity are required")
+	}
+	semaphore := make(chan struct{}, c.Capacity)
+	var workers sync.WaitGroup
+	consumeContext, err := c.Consumer.Consume(func(message jetstream.Msg) {
+		select {
+		case semaphore <- struct{}{}:
+		case <-ctx.Done():
+			return
+		}
+		workers.Add(1)
+		go func() {
+			defer workers.Done()
+			defer func() { <-semaphore }()
+			if err := c.Processor.ProcessUntilDone(ctx, message); err != nil && !errors.Is(err, context.Canceled) && c.OnError != nil {
+				c.OnError(err)
+			}
+		}()
+	}, jetstream.PullMaxMessages(c.Capacity))
+	if err != nil {
+		return fmt.Errorf("start JetStream consumer: %w", err)
+	}
+
+	select {
+	case <-ctx.Done():
+		consumeContext.Stop()
+		<-consumeContext.Closed()
+		workers.Wait()
+		return nil
+	case <-consumeContext.Closed():
+		workers.Wait()
+		return errors.New("JetStream consumer stopped unexpectedly")
+	}
+}
diff --git a/internal/assignmentqueue/jetstream_test.go b/internal/assignmentqueue/jetstream_test.go
index 13619de..cdb4f22 100644
--- a/internal/assignmentqueue/jetstream_test.go
+++ b/internal/assignmentqueue/jetstream_test.go
@@ -53,11 +53,16 @@ func TestPublisherUsesBackendSubjectAndAssignmentDeduplication(t *testing.T) {
 }
 
 type fakeHandler struct {
-	done bool
-	err  error
+	done      bool
+	err       error
+	remaining int
 }
 
-func (h fakeHandler) Handle(context.Context, taskassignment.Assignment) (bool, error) {
+func (h *fakeHandler) Handle(context.Context, taskassignment.Assignment) (bool, error) {
+	if h.remaining > 0 {
+		h.remaining--
+		return false, nil
+	}
 	return h.done, h.err
 }
 
@@ -86,7 +91,7 @@ func encodedAssignment(t *testing.T) []byte {
 
 func TestProcessorAcknowledgesOnlyCompletedAssignment(t *testing.T) {
 	message := &fakeMessage{data: encodedAssignment(t)}
-	processor := Processor{TrustDomain: "ddupan.top", Handler: fakeHandler{done: true}}
+	processor := Processor{TrustDomain: "ddupan.top", Handler: &fakeHandler{done: true}}
 	if err := processor.Process(context.Background(), message); err != nil {
 		t.Fatal(err)
 	}
@@ -97,7 +102,7 @@ func TestProcessorAcknowledgesOnlyCompletedAssignment(t *testing.T) {
 
 func TestProcessorKeepsRunningAssignmentPending(t *testing.T) {
 	message := &fakeMessage{data: encodedAssignment(t)}
-	processor := Processor{TrustDomain: "ddupan.top", Handler: fakeHandler{}}
+	processor := Processor{TrustDomain: "ddupan.top", Handler: &fakeHandler{}}
 	if err := processor.Process(context.Background(), message); err != nil {
 		t.Fatal(err)
 	}
@@ -110,7 +115,7 @@ func TestProcessorRetriesBackendFailureAndTerminatesPoisonMessage(t *testing.T)
 	retry := &fakeMessage{data: encodedAssignment(t)}
 	processor := Processor{
 		TrustDomain: "ddupan.top",
-		Handler:     fakeHandler{err: errors.New("backend unavailable")},
+		Handler:     &fakeHandler{err: errors.New("backend unavailable")},
 		RetryDelay:  time.Minute,
 	}
 	if err := processor.Process(context.Background(), retry); err == nil {
@@ -128,3 +133,52 @@ func TestProcessorRetriesBackendFailureAndTerminatesPoisonMessage(t *testing.T)
 		t.Fatalf("poison message = %#v", poison)
 	}
 }
+
+func TestProcessUntilDoneReconcilesWithoutRedelivery(t *testing.T) {
+	message := &fakeMessage{data: encodedAssignment(t)}
+	processor := Processor{
+		TrustDomain:  "ddupan.top",
+		Handler:      &fakeHandler{done: true, remaining: 2},
+		PollInterval: time.Millisecond,
+	}
+	if err := processor.ProcessUntilDone(context.Background(), message); err != nil {
+		t.Fatal(err)
+	}
+	if message.inProgress != 2 || message.acked != 1 || message.nacked != 0 {
+		t.Fatalf("message = %#v", message)
+	}
+}
+
+func TestProcessUntilDoneLeavesMessagePendingOnShutdown(t *testing.T) {
+	message := &fakeMessage{data: encodedAssignment(t)}
+	processor := Processor{
+		TrustDomain:  "ddupan.top",
+		Handler:      &fakeHandler{},
+		PollInterval: time.Hour,
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	if err := processor.ProcessUntilDone(ctx, message); !errors.Is(err, context.Canceled) {
+		t.Fatalf("error = %v", err)
+	}
+	if message.acked != 0 || message.nacked != 0 || message.terminated != 0 {
+		t.Fatalf("message = %#v", message)
+	}
+}
+
+type fakeConsumerManager struct{ config jetstream.ConsumerConfig }
+
+func (m *fakeConsumerManager) CreateOrUpdateConsumer(_ context.Context, _ string, config jetstream.ConsumerConfig) (jetstream.Consumer, error) {
+	m.config = config
+	return nil, nil
+}
+
+func TestOpenConsumerUsesIndependentDurablePerBackend(t *testing.T) {
+	manager := &fakeConsumerManager{}
+	if _, err := OpenConsumer(context.Background(), manager, "CI_RUNNER", "ci.assignment", taskassignment.BackendPod, 4); err != nil {
+		t.Fatal(err)
+	}
+	if manager.config.Durable != "pod" || manager.config.FilterSubject != "ci.assignment.pod" || manager.config.AckPolicy != jetstream.AckExplicitPolicy || manager.config.MaxAckPending != 4 {
+		t.Fatalf("config = %#v", manager.config)
+	}
+}