diff --git a/docs/PRE-RELEASE-TASKS.md b/docs/PRE-RELEASE-TASKS.md
index 53f8b58..43c9136 100644
--- a/docs/PRE-RELEASE-TASKS.md
+++ b/docs/PRE-RELEASE-TASKS.md
@@ -98,7 +98,7 @@ These land together because both edit `domain/job.go` and `storage/store.go`.
 - [x] P3.2 — Split dispatch into `app/run.go`
 - [x] P3.3 — Rework `RunDue`/`executeRun` for mode + overlap policy
 - [x] P3.4 — Settings Queue selects
-- [ ] P3.5 — Queue tests
+- [x] P3.5 — Queue tests
 
 ### Phase 4 — Notifications + Command browse
 - [ ] P4.1 — Failure notifications
diff --git a/src/app/run_test.go b/src/app/run_test.go
new file mode 100644
index 0000000..b8feb5f
--- /dev/null
+++ b/src/app/run_test.go
@@ -0,0 +1,288 @@
+package app
+
+import (
+	"context"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"gitea.mixdep.ru/mix/gosentry/src/domain"
+)
+
+// newQueueService builds a temp-backed Service with a chosen execution mode and
+// overlap policy. Config is set before any run starts, so mutating it directly is
+// safe (construction is single-threaded).
+func newQueueService(t *testing.T, mode domain.ExecutionMode, policy domain.OverlapPolicy, jobs []domain.Job) *Service {
+	t.Helper()
+	svc := newTempService(t, jobs)
+	svc.store.Config.ExecutionMode = mode
+	svc.store.Config.OverlapPolicy = policy
+	return svc
+}
+
+// primeDue forces the given jobs to be due by backdating their NextDue. Tests tick
+// RunDue at the real wall clock; a long "@every 1h" schedule then advances a
+// started job's NextDue an hour out, so it does not spuriously re-fire on a later
+// tick unless a test re-primes it.
+func primeDue(t *testing.T, svc *Service, ids ...int) {
+	t.Helper()
+	svc.mu.Lock()
+	defer svc.mu.Unlock()
+	for _, id := range ids {
+		if rt := svc.runtimes[id]; rt != nil {
+			rt.NextDue = time.Now().Add(-time.Second)
+		}
+	}
+}
+
+// completions subscribes a recorder that forwards every RunRecorded onto a
+// channel, so tests can wait for runs to finish (and drain in-flight runs before
+// the temp dir is cleaned up).
+func completions(svc *Service) <-chan domain.RunRecord {
+	ch := make(chan domain.RunRecord, 16)
+	svc.Subscribe(ObserverFunc(func(e Event) {
+		if rr, ok := e.(RunRecorded); ok {
+			ch <- rr.Record
+		}
+	}))
+	return ch
+}
+
+func waitRecord(t *testing.T, ch <-chan domain.RunRecord) domain.RunRecord {
+	t.Helper()
+	select {
+	case r := <-ch:
+		return r
+	case <-time.After(2 * time.Second):
+		t.Fatal("timed out waiting for a run record")
+		return domain.RunRecord{}
+	}
+}
+
+func expectNoEntry(t *testing.T, entered <-chan int) {
+	t.Helper()
+	select {
+	case id := <-entered:
+		t.Fatalf("a job (%d) started unexpectedly", id)
+	case <-time.After(200 * time.Millisecond):
+	}
+}
+
+// TestRunDueParallelStartsAllDueJobs verifies that in parallel mode every due job
+// starts at once: both runs are in flight (blocked in the runner) before either
+// is released.
+func TestRunDueParallelStartsAllDueJobs(t *testing.T) {
+	svc := newQueueService(t, domain.ExecutionModeParallel, domain.OverlapPolicySkip, []domain.Job{
+		{ID: 1, Name: "A", Schedule: "@every 1h", Command: "echo", Enabled: true},
+		{ID: 2, Name: "B", Schedule: "@every 1h", Command: "echo", Enabled: true},
+	})
+
+	entered := make(chan int, 2)
+	release := make(chan struct{})
+	svc.runJob = func(_ context.Context, job *domain.Job, _ string, _ string) domain.RunRecord {
+		entered <- job.ID
+		<-release
+		return domain.RunRecord{Time: "t", JobID: job.ID, JobName: job.Name, State: "Success"}
+	}
+	done := completions(svc)
+
+	primeDue(t, svc, 1, 2)
+	svc.RunDue(time.Now())
+
+	// Both runs must reach the runner (and block) before any is released, which is
+	// only possible if RunDue started them concurrently.
+	got := map[int]bool{}
+	for i := 0; i < 2; i++ {
+		select {
+		case id := <-entered:
+			got[id] = true
+		case <-time.After(2 * time.Second):
+			t.Fatalf("only %d job(s) started in parallel, want 2", len(got))
+		}
+	}
+	if !got[1] || !got[2] {
+		t.Fatalf("started jobs = %v, want both 1 and 2", got)
+	}
+
+	close(release)
+	waitRecord(t, done)
+	waitRecord(t, done)
+}
+
+// TestRunDueSequentialSerializes verifies that in sequential mode only one due job
+// runs at a time: the second due job waits until the first finishes and a later
+// tick picks it up.
+func TestRunDueSequentialSerializes(t *testing.T) {
+	svc := newQueueService(t, domain.ExecutionModeSequential, domain.OverlapPolicySkip, []domain.Job{
+		{ID: 1, Name: "A", Schedule: "@every 1h", Command: "echo", Enabled: true},
+		{ID: 2, Name: "B", Schedule: "@every 1h", Command: "echo", Enabled: true},
+	})
+
+	entered := make(chan int, 2)
+	release := make(chan struct{})
+	svc.runJob = func(_ context.Context, job *domain.Job, _ string, _ string) domain.RunRecord {
+		entered <- job.ID
+		<-release
+		return domain.RunRecord{Time: "t", JobID: job.ID, JobName: job.Name, State: "Success"}
+	}
+	done := completions(svc)
+
+	primeDue(t, svc, 1, 2)
+	svc.RunDue(time.Now())
+
+	// Exactly one job — the first in order — starts; the second is held back.
+	if id := <-entered; id != 1 {
+		t.Fatalf("first started job = %d, want 1", id)
+	}
+	expectNoEntry(t, entered)
+
+	// Let the first run finish, then tick again: now the second job runs.
+	close(release)
+	waitRecord(t, done)
+	svc.RunDue(time.Now())
+	if id := <-entered; id != 2 {
+		t.Fatalf("second started job = %d, want 2", id)
+	}
+	waitRecord(t, done)
+}
+
+// TestRunDueSkipDropsOverlap verifies that under the "skip" overlap policy a job
+// coming due again while its own run is in flight does not queue or start a second
+// run.
+func TestRunDueSkipDropsOverlap(t *testing.T) {
+	svc := newQueueService(t, domain.ExecutionModeParallel, domain.OverlapPolicySkip, []domain.Job{
+		{ID: 1, Name: "A", Schedule: "@every 1h", Command: "echo", Enabled: true},
+	})
+
+	entered := make(chan int, 2)
+	release := make(chan struct{})
+	var calls int32
+	svc.runJob = func(_ context.Context, job *domain.Job, _ string, _ string) domain.RunRecord {
+		atomic.AddInt32(&calls, 1)
+		entered <- job.ID
+		<-release
+		return domain.RunRecord{Time: "t", JobID: job.ID, JobName: job.Name, State: "Success"}
+	}
+	done := completions(svc)
+
+	primeDue(t, svc, 1)
+	svc.RunDue(time.Now())
+	if id := <-entered; id != 1 {
+		t.Fatalf("started job = %d, want 1", id)
+	}
+
+	// The job is now in flight; make it due again and tick. Skip must drop it.
+	primeDue(t, svc, 1)
+	svc.RunDue(time.Now())
+	expectNoEntry(t, entered)
+
+	svc.mu.Lock()
+	pending := svc.runtimes[1].Pending
+	svc.mu.Unlock()
+	if pending {
+		t.Error("skip policy must not mark the job Pending")
+	}
+
+	close(release)
+	waitRecord(t, done)
+	// No re-run is queued, so the runner is invoked exactly once.
+	expectNoEntry(t, entered)
+	if got := atomic.LoadInt32(&calls); got != 1 {
+		t.Errorf("runner called %d time(s), want 1", got)
+	}
+}
+
+// TestRunDueQueueRerunsAfterFinish verifies that under the "queue" overlap policy a
+// job coming due again while running is marked Pending and re-run as soon as the
+// in-flight run finishes.
+func TestRunDueQueueRerunsAfterFinish(t *testing.T) {
+	svc := newQueueService(t, domain.ExecutionModeParallel, domain.OverlapPolicyQueue, []domain.Job{
+		{ID: 1, Name: "A", Schedule: "@every 1h", Command: "echo", Enabled: true},
+	})
+
+	entered := make(chan int, 2)
+	release := make(chan struct{})
+	var calls int32
+	svc.runJob = func(_ context.Context, job *domain.Job, _ string, _ string) domain.RunRecord {
+		atomic.AddInt32(&calls, 1)
+		entered <- job.ID
+		<-release
+		return domain.RunRecord{Time: "t", JobID: job.ID, JobName: job.Name, State: "Success"}
+	}
+	done := completions(svc)
+
+	primeDue(t, svc, 1)
+	svc.RunDue(time.Now())
+	if id := <-entered; id != 1 {
+		t.Fatalf("started job = %d, want 1", id)
+	}
+
+	// Re-due the running job and tick: queue must mark it Pending without starting
+	// a second concurrent run.
+	primeDue(t, svc, 1)
+	svc.RunDue(time.Now())
+	expectNoEntry(t, entered)
+
+	svc.mu.Lock()
+	pending := svc.runtimes[1].Pending
+	svc.mu.Unlock()
+	if !pending {
+		t.Fatal("queue policy must mark the job Pending")
+	}
+
+	// Releasing the first run lets executeRun start the deferred run automatically.
+	close(release)
+	waitRecord(t, done)
+	if id := <-entered; id != 1 {
+		t.Fatalf("re-run job = %d, want 1", id)
+	}
+	waitRecord(t, done)
+
+	if got := atomic.LoadInt32(&calls); got != 2 {
+		t.Errorf("runner called %d time(s), want 2 (original + queued re-run)", got)
+	}
+	svc.mu.Lock()
+	pending = svc.runtimes[1].Pending
+	svc.mu.Unlock()
+	if pending {
+		t.Error("Pending must be cleared after the re-run starts")
+	}
+}
+
+// TestRunNowSequentialGuard verifies the sequential-mode guard in RunNow: a manual
+// run is refused while another job is running, and allowed once nothing is.
+func TestRunNowSequentialGuard(t *testing.T) {
+	svc := newQueueService(t, domain.ExecutionModeSequential, domain.OverlapPolicySkip, []domain.Job{
+		{ID: 1, Name: "A", Schedule: "@every 1h", Command: "echo", Enabled: true},
+		{ID: 2, Name: "B", Schedule: "@every 1h", Command: "echo", Enabled: true},
+	})
+
+	entered := make(chan int, 2)
+	release := make(chan struct{})
+	svc.runJob = func(_ context.Context, job *domain.Job, _ string, _ string) domain.RunRecord {
+		entered <- job.ID
+		<-release
+		return domain.RunRecord{Time: "t", JobID: job.ID, JobName: job.Name, State: "Success"}
+	}
+	done := completions(svc)
+
+	if err := svc.RunNow(1); err != nil {
+		t.Fatalf("first RunNow: %v", err)
+	}
+	<-entered // job 1 is in flight
+
+	if err := svc.RunNow(2); err == nil {
+		t.Error("expected RunNow to be refused while another job runs (sequential mode)")
+	}
+
+	// Once job 1 finishes, a manual run of job 2 is allowed.
+	close(release)
+	waitRecord(t, done)
+	if err := svc.RunNow(2); err != nil {
+		t.Fatalf("RunNow after first finished: %v", err)
+	}
+	if id := <-entered; id != 2 {
+		t.Fatalf("second manual run job = %d, want 2", id)
+	}
+	waitRecord(t, done)
+}