test✅: fail on an Append error this test does not model

Only ErrQueueClosed was examined and every other error was discarded, so a
run where nothing reached a queue at all would leave the refusal count at
zero and the test green. The first unexpected error is now kept and fails
the test.

The pool is sized so a full queue cannot be one of them: it returns an error
of its own and is expected while the consumer is held, which would otherwise
make the new check fire on the normal path.

.github/workflows/go.yml

@@ -15,6 +15,27 @@ jobs:
   build:
     name: Build
     runs-on: ubuntu-latest
+
+    # The queue's ordering rule - consumers registered before the queue is
+    # started - is invisible on the memory backend, which is the default and
+    # therefore what every other test runs on: queue.Memory's Register starts a
+    # consumer goroutine whatever the state. Only redis refuses a late
+    # registration, so without a server here the tests that cover it would skip
+    # and the suite would report success for a queue that accepts no consumers.
+    services:
+      redis:
+        image: redis:7-alpine
+        ports:
+          - 6379:6379
+        options: >-
+          --health-cmd "redis-cli ping"
+          --health-interval 5s
+          --health-timeout 3s
+          --health-retries 10
+
+    env:
+      GO_ADMIN_TEST_REDIS_ADDR: 127.0.0.1:6379
+
     steps:
 
     - name: Set up Go 1.26

app/jobs/jobbase.go

@@ -1,6 +1,7 @@
 package jobs
 
 import (
+	"context"
 	"fmt"
 	log "github.com/go-admin-team/go-admin-core/v2/logger"
 	"github.com/go-admin-team/go-admin-core/v2/sdk"
@@ -145,11 +146,36 @@ func setup(key string, db *gorm.DB) {
 	}
 
 	// 其中任务
-	crontab.Start()
+	startCrontab(crontab)
+}
+
+// startCrontab starts c and arranges for it to be stopped on the way out.
+//
+// The stop used to be `defer crontab.Stop()` followed by `select {}`. The
+// select never returned, so the defer never ran and the scheduler was never
+// stopped; and because setup never returned, the loop in Setup never reached
+// the second tenant - only whichever database came first out of the map ever
+// got a scheduler at all. cron.Start is itself `go c.run()`, so the select was
+// blocking for nothing.
+//
+// cron.Stop returns a context that closes once the jobs already running have
+// finished. That is the wait the shutdown budget exists to bound: giving up on
+// it leaves those jobs running until the process exits, which is better than
+// holding the whole shutdown open for one job that will not end.
+func startCrontab(c *cron.Cron) {
+	c.Start()
 	fmt.Println(time.Now().Format(timeFormat), " [INFO] JobCore start success.")
+
 	// 关闭任务
-	defer crontab.Stop()
-	select {}
+	sdk.Runtime.SetShutdown(func(ctx context.Context) {
+		stopped := c.Stop()
+		select {
+		case <-stopped.Done():
+			fmt.Println(time.Now().Format(timeFormat), " [INFO] JobCore stopped.")
+		case <-ctx.Done():
+			fmt.Println(time.Now().Format(timeFormat), " [WARN] JobCore stop gave up waiting for running jobs")
+		}
+	})
 }
 
 // AddJob 添加任务 AddJob(invokeTarget string, jobId int, jobName string, cronExpression string)

app/jobs/jobbase_test.go

@@ -0,0 +1,52 @@
+package jobs
+
+import (
+	"context"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/go-admin-team/go-admin-core/v2/sdk"
+	"github.com/go-admin-team/go-admin-core/v2/sdk/pkg/cronjob"
+)
+
+// The scheduler had never been stopped. `defer crontab.Stop()` sat directly
+// above a `select {}` that never returned, so the deferred call was
+// unreachable for the life of the process.
+//
+// There is one test rather than several because BeforeExit closes to further
+// registration once it has run: a second RunShutdown in this binary would find
+// an empty registry and pass while proving nothing.
+func TestTheSchedulerIsStoppedOnTheWayOut(t *testing.T) {
+	var ticks atomic.Int64
+
+	c := cronjob.NewWithSeconds()
+	if _, err := c.AddFunc("* * * * * *", func() { ticks.Add(1) }); err != nil {
+		t.Fatalf("AddFunc: %v", err)
+	}
+
+	startCrontab(c)
+
+	// It has to be running before stopping it can mean anything.
+	deadline := time.Now().Add(5 * time.Second)
+	for ticks.Load() == 0 && time.Now().Before(deadline) {
+		time.Sleep(20 * time.Millisecond)
+	}
+	if ticks.Load() == 0 {
+		t.Fatal("the scheduler never ran the job, so this test cannot show it was stopped")
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Second)
+	defer cancel()
+	if err := sdk.Runtime.RunShutdown(ctx); err != nil {
+		t.Fatalf("RunShutdown: %v", err)
+	}
+
+	// Two and a half seconds is two more firings of a job that runs every
+	// second, so silence here is the assertion.
+	at := ticks.Load()
+	time.Sleep(2500 * time.Millisecond)
+	if n := ticks.Load() - at; n > 0 {
+		t.Errorf("the job fired %d more times after shutdown: the scheduler is still running", n)
+	}
+}

app/other/router/monitor.go

@@ -1,23 +1,63 @@
 package router
 
 import (
+	"context"
 	"net/http"
+	"time"
 
 	"github.com/gin-gonic/gin"
 	"github.com/go-admin-team/go-admin-core/v2/tools/transfer"
 	"github.com/prometheus/client_golang/prometheus/promhttp"
+
+	"go-admin/common/health"
 )
 
 func init() {
 	routerNoCheckRole = append(routerNoCheckRole, registerMonitorRouter)
 }
 
-// 需认证的路由代码
+// readyTimeout bounds the whole probe. It has to stay under whatever period
+// the orchestrator polls on, or a slow dependency turns a readiness check into
+// a queue of readiness checks.
+const readyTimeout = 2 * time.Second
+
+// 无需认证的路由代码
 func registerMonitorRouter(v1 *gin.RouterGroup) {
 	v1.GET("/metrics", transfer.Handler(promhttp.Handler()))
-	//健康检查
+
+	// 健康检查（存活）
+	//
+	// Stays a bare 200 on purpose. This is the answer to "should I restart
+	// you", and a process whose database is unreachable does not want
+	// restarting - that turns one outage into a crash loop and throws away the
+	// connection pool, the cache and every in-flight request along the way.
 	v1.GET("/health", func(c *gin.Context) {
 		c.Status(http.StatusOK)
 	})
 
-}
+	// 就绪检查
+	//
+	// The answer to "should I send you requests". It fails while a dependency
+	// is unreachable, and from the moment shutdown begins - which is before
+	// the server stops accepting, so a load balancer can take this instance
+	// out of the pool while it can still finish what it has.
+	v1.GET("/ready", func(c *gin.Context) {
+		if health.Draining() {
+			c.JSON(http.StatusServiceUnavailable, gin.H{
+				"status": "draining",
+				"checks": []health.Check{},
+			})
+			return
+		}
+		ctx, cancel := context.WithTimeout(c.Request.Context(), readyTimeout)
+		defer cancel()
+
+		checks := health.Ready(ctx)
+		status := http.StatusOK
+		if !health.Healthy(checks) {
+			status = http.StatusServiceUnavailable
+		}
+		c.JSON(status, gin.H{"status": http.StatusText(status), "checks": checks})
+	})
+
+}

cmd/api/phase_test.go

@@ -0,0 +1,160 @@
+package api
+
+import (
+	"fmt"
+	"net"
+	"net/http"
+	"testing"
+	"time"
+
+	"github.com/go-admin-team/go-admin-core/v2/sdk"
+	"github.com/go-admin-team/go-admin-core/v2/sdk/config"
+	"github.com/go-admin-team/go-admin-core/v2/sdk/runtime"
+)
+
+// freePort returns a port nothing is listening on. It is inherently a guess -
+// the port is free when it is handed back and could be taken a moment later -
+// but every alternative needs the caller to hold the listener, which is the one
+// thing these tests cannot do.
+func freePort(t *testing.T) int {
+	t.Helper()
+	ln, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		t.Fatalf("probe listen: %v", err)
+	}
+	port := ln.Addr().(*net.TCPAddr).Port
+	_ = ln.Close()
+	return port
+}
+
+// AfterListen promises a hook that the port is reachable. Both halves of that
+// are asserted here, and in one test rather than two, because the phase seals
+// itself once it has run: a second test calling RunPhase again would find a
+// closed registry and pass while proving nothing.
+//
+// The failing bind comes first for the same reason. It must leave the phase
+// unsealed, which is only visible if nothing has sealed it yet.
+func TestAfterListenIsAnnouncedOnlyOnceThePortIsBound(t *testing.T) {
+	// The pause makes the "announced synchronously" claim testable: if the
+	// announcement were moved onto a goroutine, startServing would return
+	// while the hook was still sleeping and the count below would be zero.
+	var ran int
+	sdk.Runtime.SetPhase(runtime.AfterListen, func() {
+		time.Sleep(50 * time.Millisecond)
+		ran++
+	})
+
+	// Somebody else already has the port. Under ListenAndServe this surfaced
+	// on the serving goroutine, far too late to stop the announcement.
+	taken, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		t.Fatalf("occupy: %v", err)
+	}
+	defer func() { _ = taken.Close() }()
+
+	blocked := &http.Server{Addr: taken.Addr().String(), Handler: http.NewServeMux()}
+	if err := startServing(blocked, false, "", ""); err == nil {
+		t.Fatal("startServing returned no error for a port that was already taken")
+	}
+	if ran != 0 {
+		t.Errorf("AfterListen ran %d times after a failed bind; a hook there is told the port is reachable", ran)
+	}
+	if sdk.Runtime.PhaseSealed(runtime.AfterListen) {
+		t.Error("a failed bind sealed AfterListen, so the phase could never run for a server that did start")
+	}
+
+	// A certificate that cannot be read is the other way to fail before there
+	// is anything to announce. ServeTLS reads it on the serving goroutine, so
+	// without the check in startServing this would be a hook told the port was
+	// reachable while the server was already on its way down.
+	if err := startServing(&http.Server{Addr: "127.0.0.1:0"}, true, "no-such.pem", "no-such.key"); err == nil {
+		t.Fatal("startServing returned no error for a certificate that does not exist")
+	}
+	if ran != 0 {
+		t.Errorf("AfterListen ran %d times after a certificate failure", ran)
+	}
+	if sdk.Runtime.PhaseSealed(runtime.AfterListen) {
+		t.Error("a certificate failure sealed AfterListen")
+	}
+
+	// And now a bind that works.
+	port := freePort(t)
+	srv := &http.Server{Addr: fmt.Sprintf("127.0.0.1:%d", port), Handler: http.NewServeMux()}
+	if err := startServing(srv, false, "", ""); err != nil {
+		t.Fatalf("startServing on a free port: %v", err)
+	}
+	defer func() { _ = srv.Close() }()
+
+	// Checked the instant startServing returns, so this is also the assertion
+	// that it did not return early: an asynchronous announcement would still
+	// be inside the sleep. Synchrony matters because an announcement that
+	// overlaps the wait below could, on a fast SIGTERM, have the shutdown
+	// callbacks finish before the startup ones.
+	if ran != 1 {
+		t.Fatalf("AfterListen ran %d times, want 1", ran)
+	}
+
+	// The claim is not "Serve was called" but "the port answers". Dial it.
+	c, err := net.DialTimeout("tcp", srv.Addr, 5*time.Second)
+	if err != nil {
+		t.Fatalf("AfterListen ran but the port does not answer: %v", err)
+	}
+	_ = c.Close()
+}
+
+// BeforeRouter is the last point at which a module can still affect how routes
+// are built, so it has to run while there is no engine yet. The before registry
+// is a different moment despite the name: those callbacks run after initRouter
+// has built the engine.
+//
+// The two are two lines apart in buildRouter, and calling them equivalent is a
+// mistake this repository has already made in writing. Until this test the
+// ordering was checked by reading - which is how the stop signals came to be
+// armed after the readiness banner in the same file.
+func TestBeforeRouterRunsWhileThereIsNoEngine(t *testing.T) {
+	freshRuntime(t)
+
+	// AuthInit reads these two package-level values and nothing else. No
+	// database is involved in building a router: the handlers are registered,
+	// not called.
+	config.ApplicationConfig.Mode = "dev"
+	config.JwtConfig.Secret = "test-secret-for-the-router-build"
+
+	type observation struct {
+		ran        int
+		engineWas  interface{}
+		engineSeen bool
+	}
+	var phase, before observation
+
+	sdk.Runtime.SetPhase(runtime.BeforeRouter, func() {
+		phase.ran++
+		phase.engineWas = sdk.Runtime.GetEngine()
+		phase.engineSeen = true
+	})
+	sdk.Runtime.SetBefore(func() {
+		before.ran++
+		before.engineWas = sdk.Runtime.GetEngine()
+		before.engineSeen = true
+	})
+
+	buildRouter()
+
+	if phase.ran != 1 {
+		t.Fatalf("BeforeRouter ran %d times, want 1", phase.ran)
+	}
+	if !phase.engineSeen || phase.engineWas != nil {
+		t.Errorf("BeforeRouter saw engine %v, want nil: it is meant to run before initRouter builds one", phase.engineWas)
+	}
+
+	if before.ran != 1 {
+		t.Fatalf("the before registry ran %d times, want 1", before.ran)
+	}
+	if before.engineWas == nil {
+		t.Error("a before callback saw no engine; that registry is meant to run after initRouter, and describing it as equivalent to BeforeRouter is the error this asserts against")
+	}
+
+	if sdk.Runtime.GetEngine() == nil {
+		t.Error("buildRouter returned with no engine built")
+	}
+}

cmd/api/queue_test.go

@@ -0,0 +1,167 @@
+package api
+
+import (
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	corestorage "github.com/go-admin-team/go-admin-core/v2/storage"
+)
+
+// recordingQueue records what was done to it, in order. Register and Run are
+// the two calls whose order is the point of this file; Append and Shutdown are
+// here to satisfy the interface.
+type recordingQueue struct {
+	mu     sync.Mutex
+	events []string
+	ran    chan struct{}
+}
+
+func newRecordingQueue() *recordingQueue {
+	return &recordingQueue{ran: make(chan struct{}, 4)}
+}
+
+func (q *recordingQueue) record(e string) {
+	q.mu.Lock()
+	q.events = append(q.events, e)
+	q.mu.Unlock()
+}
+
+func (q *recordingQueue) seen() []string {
+	q.mu.Lock()
+	defer q.mu.Unlock()
+	return append([]string(nil), q.events...)
+}
+
+func (q *recordingQueue) String() string                    { return "recording" }
+func (q *recordingQueue) Append(corestorage.Messager) error { return nil }
+func (q *recordingQueue) Register(name string, _ corestorage.ConsumerFunc) {
+	q.record("register:" + name)
+}
+func (q *recordingQueue) Shutdown() {}
+
+func (q *recordingQueue) Run() {
+	q.record("run")
+	select {
+	case q.ran <- struct{}{}:
+	default:
+	}
+}
+
+// waitForRun waits for Run, which is started on a goroutine.
+func (q *recordingQueue) waitForRun(t *testing.T) {
+	t.Helper()
+	select {
+	case <-q.ran:
+	case <-time.After(5 * time.Second):
+		t.Fatalf("Run was never called; saw: %v", q.seen())
+	}
+}
+
+// The consumers must be registered before the queue is started. A queue that
+// is already running refuses further registration - the contract
+// implementations answer storage.ErrQueueAlreadyStarted - and the legacy
+// adapter this path goes through drops that error, so the wrong order loses
+// consumers with nothing said about it. The memory backend does not care,
+// which is exactly why this cannot be left to be noticed in use.
+func TestConsumersAreRegisteredBeforeTheQueueIsStarted(t *testing.T) {
+	attachedQueue.Store(0)
+	t.Cleanup(func() { attachedQueue.Store(0) })
+
+	q := newRecordingQueue()
+	attachConsumersOnce(1, q)
+	q.waitForRun(t)
+
+	seen := q.seen()
+	runAt := -1
+	registers := 0
+	for i, e := range seen {
+		switch {
+		case e == "run":
+			if runAt < 0 {
+				runAt = i
+			}
+		case strings.HasPrefix(e, "register:"):
+			registers++
+			if runAt >= 0 {
+				t.Errorf("%q came after Run; a running queue refuses registration", e)
+			}
+		}
+	}
+	if registers != 3 {
+		t.Errorf("registered %d consumers, want 3; saw %v", registers, seen)
+	}
+	if runAt < 0 {
+		t.Errorf("the queue was never started; saw %v", seen)
+	}
+}
+
+// AfterResource runs again on every configuration reload, so the hook has to
+// be idempotent with respect to a given queue - not "does nothing the second
+// time". Registering twice on the same queue would give every message two
+// consumers and write every log row twice.
+func TestTheSameQueueIsNotGivenConsumersTwice(t *testing.T) {
+	attachedQueue.Store(0)
+	t.Cleanup(func() { attachedQueue.Store(0) })
+
+	q := newRecordingQueue()
+	attachConsumersOnce(1, q)
+	q.waitForRun(t)
+	attachConsumersOnce(1, q)
+
+	// Nothing to wait for on the second call, so give a wrong implementation
+	// the time it would need to show up.
+	time.Sleep(200 * time.Millisecond)
+	if n := len(q.seen()); n != 4 {
+		t.Errorf("%d calls after attaching twice to the same queue, want 4 (3 registers + 1 run); saw %v", n, q.seen())
+	}
+}
+
+// The other half of the same rule: a reload builds a new adapter, and the
+// consumers on the old one are attached to a queue nobody publishes to any
+// more. A new generation must get its own set.
+func TestANewQueueGetsItsOwnConsumers(t *testing.T) {
+	attachedQueue.Store(0)
+	t.Cleanup(func() { attachedQueue.Store(0) })
+
+	first := newRecordingQueue()
+	attachConsumersOnce(1, first)
+	first.waitForRun(t)
+
+	second := newRecordingQueue()
+	attachConsumersOnce(2, second)
+	second.waitForRun(t)
+
+	if n := len(second.seen()); n != 4 {
+		t.Errorf("the queue from the second generation saw %d calls, want 4; saw %v", n, second.seen())
+	}
+	if n := len(first.seen()); n != 4 {
+		t.Errorf("the queue from the first generation saw %d calls, want 4 - it should not have been touched again; saw %v", n, first.seen())
+	}
+}
+
+// Generation 0 means the configuration has no queue section at all, so nothing
+// was installed and the runtime hands back its own memory queue. That case
+// still has to get consumers - the registration it replaces was unconditional,
+// and dropping it would stop the login and operation logs for anyone who
+// commented the section out.
+func TestAnUnconfiguredQueueStillGetsConsumers(t *testing.T) {
+	attachedQueue.Store(0)
+	t.Cleanup(func() { attachedQueue.Store(0) })
+
+	q := newRecordingQueue()
+	attachConsumersOnce(0, q)
+	q.waitForRun(t)
+
+	if n := len(q.seen()); n != 4 {
+		t.Errorf("an unconfigured queue saw %d calls, want 4; saw %v", n, q.seen())
+	}
+
+	// And still only once.
+	attachConsumersOnce(0, q)
+	time.Sleep(200 * time.Millisecond)
+	if n := len(q.seen()); n != 4 {
+		t.Errorf("generation 0 was attached to twice: %d calls, want 4; saw %v", n, q.seen())
+	}
+}

cmd/api/server.go

@@ -2,10 +2,13 @@ package api
 
 import (
 	"context"
+	"crypto/tls"
 	"fmt"
+	"net"
 	"net/http"
 	"os"
 	"os/signal"
+	"sync/atomic"
 	"syscall"
 	"time"
 
@@ -14,8 +17,11 @@ import (
 	log "github.com/go-admin-team/go-admin-core/v2/logger"
 	"github.com/go-admin-team/go-admin-core/v2/sdk"
 	"github.com/go-admin-team/go-admin-core/v2/sdk/api"
+	"github.com/go-admin-team/go-admin-core/v2/sdk/bootstrap"
 	"github.com/go-admin-team/go-admin-core/v2/sdk/config"
 	"github.com/go-admin-team/go-admin-core/v2/sdk/pkg"
+	"github.com/go-admin-team/go-admin-core/v2/sdk/runtime"
+	corestorage "github.com/go-admin-team/go-admin-core/v2/storage"
 	"github.com/pkg/errors"
 	"github.com/spf13/cobra"
 
@@ -24,6 +30,7 @@ import (
 	"go-admin/app/jobs"
 	"go-admin/common/database"
 	"go-admin/common/global"
+	"go-admin/common/health"
 	common "go-admin/common/middleware"
 	"go-admin/common/middleware/handler"
 	"go-admin/common/storage"
@@ -60,30 +67,101 @@ func init() {
 func setup() {
 	// 注入配置扩展项
 	config.ExtendConfig = &ext.ExtConfig
+
+	// Registered before the configuration is read. SetupConfig announces
+	// AfterResource as soon as the callbacks that build the resources have
+	// run, so a hook added after that call would miss the first round and the
+	// queue would have no consumers until somebody edited the config file.
+	sdk.Runtime.SetPhase(runtime.AfterResource, attachQueueConsumers)
+
+	// On AfterListen rather than on a bare goroutine from run(). Two reasons:
+	// the phase runs behind core's panic guard, which does not reach across a
+	// goroutine boundary - a panic while loading jobs used to take the whole
+	// process down with a stack that named this file - and the jobs it starts
+	// can call the API, which is only true once the socket is accepting.
+	sdk.Runtime.SetPhase(runtime.AfterListen, startCronJobs)
+
 	//1. 读取配置
-	config.Setup(
+	bootstrap.SetupConfig(
 		file.NewSource(file.WithPath(configYml)),
 		database.Setup,
 		storage.Setup,
 	)
-	//注册监听函数
-	queue := sdk.Runtime.GetQueuePrefix("")
-	queue.Register(global.LoginLog, models.SaveLoginLog)
-	queue.Register(global.OperateLog, models.SaveOperaLog)
-	queue.Register(global.ApiCheck, models.SaveSysApi)
-	go queue.Run()
 
 	usageStr := `starting api server...`
 	log.Info(usageStr)
 }
 
+// startCronJobs registers the job implementations and starts a scheduler for
+// every tenant database.
+//
+// It is synchronous, like the phase that runs it. jobs.Setup returns now that
+// the `select {}` at the end of its per-tenant setup is gone, which is what
+// makes that possible; while it was there this could only be a goroutine, and
+// a goroutine is outside the panic guard.
+func startCronJobs() {
+	jobs.InitJob()
+	jobs.Setup(sdk.Runtime.GetAllDb())
+}
+
+// attachedQueue is the queue generation the consumers are attached to, plus
+// one, so that the zero value means "attached to nothing yet". Written from
+// the goroutine running the phase, read from the next one - rounds never
+// overlap, but they are not the same goroutine.
+var attachedQueue atomic.Uint64
+
+// attachQueueConsumers registers the log consumers against the queue that is
+// current, and starts it.
+//
+// It runs on AfterResource, so it runs again after every configuration reload
+// - and it has to. A reload rebuilds the queue adapter, and consumers
+// registered against the one that existed at start-up are attached to an
+// adapter nobody publishes to any more, so the login and operation logs stop
+// being written with nothing said about it.
+//
+// It is therefore idempotent with respect to a given queue rather than "does
+// nothing the second time": a new adapter gets a fresh set of consumers, the
+// same one gets none. Registering twice on the same queue would give every
+// message two consumers and write every log row twice.
+//
+// Generation 0 means the configuration has no queue section, so nothing was
+// installed and GetQueuePrefix hands back the runtime's own memory queue.
+// That case still gets consumers - it is what the previous unconditional
+// registration did, and dropping it would silently stop logging for anyone who
+// commented the section out - it just never gets them twice.
+func attachQueueConsumers() {
+	attachConsumersOnce(storage.QueueGeneration(), sdk.Runtime.GetQueuePrefix(""))
+}
+
+// attachConsumersOnce puts the log consumers on q and starts it, unless gen
+// says this queue already has them.
+//
+// Split out from attachQueueConsumers so that the order and the once-ness can
+// be checked against a queue the test controls: the sequence that matters here
+// cannot be read back out of a real adapter.
+func attachConsumersOnce(gen uint64, q corestorage.AdapterQueue) {
+	if attachedQueue.Load() == gen+1 {
+		return
+	}
+	attachedQueue.Store(gen + 1)
+
+	//注册监听函数
+	q.Register(global.LoginLog, models.SaveLoginLog)
+	q.Register(global.OperateLog, models.SaveOperaLog)
+	q.Register(global.ApiCheck, models.SaveSysApi)
+
+	// Started only now, and by whoever registered. setupQueue deliberately
+	// leaves it stopped: a queue that is already running refuses further
+	// registration, and the adapter in this path drops that error on the
+	// floor, so starting first loses consumers without a word.
+	go q.Run()
+}
+
 func run() error {
 	if config.ApplicationConfig.Mode == pkg.ModeProd.String() {
 		gin.SetMode(gin.ReleaseMode)
 	}
-	initRouter()
-
-	runStartupHooks()
+	buildRouter()
 
 	srv := &http.Server{
 		Addr:         fmt.Sprintf("%s:%d", config.ApplicationConfig.Host, config.ApplicationConfig.Port),
@@ -92,12 +170,6 @@ func run() error {
 		WriteTimeout: time.Duration(config.ApplicationConfig.WriterTimeout) * time.Second,
 	}
 
-	go func() {
-		jobs.InitJob()
-		jobs.Setup(sdk.Runtime.GetAllDb())
-
-	}()
-
 	if apiCheck {
 		var routers = sdk.Runtime.GetRouter()
 		q := sdk.Runtime.GetQueuePrefix("")
@@ -122,18 +194,10 @@ func run() error {
 	// arming is separate from waiting.
 	quit, disarmStopSignals := armStopSignals()
 
-	go func() {
-		// 服务连接
-		if config.SslConfig.Enable {
-			if err := srv.ListenAndServeTLS(config.SslConfig.Pem, config.SslConfig.KeyStr); err != nil && !errors.Is(err, http.ErrServerClosed) {
-				log.Fatal("listen: ", err)
-			}
-		} else {
-			if err := srv.ListenAndServe(); err != nil && !errors.Is(err, http.ErrServerClosed) {
-				log.Fatal("listen: ", err)
-			}
-		}
-	}()
+	if err := startServing(srv, config.SslConfig.Enable, config.SslConfig.Pem, config.SslConfig.KeyStr); err != nil {
+		return err
+	}
+
 	fmt.Println(pkg.Red(string(global.LogoContent)))
 	tip()
 	fmt.Println(pkg.Green("Server run at:"))
@@ -149,6 +213,18 @@ func run() error {
 	// the default handler, so a shutdown that hangs can still be interrupted.
 	disarmStopSignals()
 
+	// Said before anything is taken apart. A configuration reload arriving in
+	// this window would otherwise re-run AfterResource - rebuilding the pool
+	// and the queue adapter, and re-registering consumers - on top of cleanup
+	// that has already run.
+	sdk.Runtime.BeginShutdown()
+	// Readiness fails from here, which is before the server stops accepting.
+	// The order is the whole point: a load balancer that is told "not ready"
+	// while this instance can still finish what it has in flight takes it out
+	// of the pool without dropping anything. Reversed, the connections are cut
+	// first and the health check reports it afterwards.
+	health.BeginDraining()
+
 	log.Info("Shutdown Server ... ")
 	if err := shutdownServer(srv, shutdownTimeout); err != nil {
 		// Not log.Fatal: that is an unconditional os.Exit(1), and Shutdown
@@ -156,15 +232,28 @@ func run() error {
 		// which is when the cleanup that follows matters most.
 		log.Error("Server Shutdown: ", err)
 	}
+
+	// Runs whether or not the line above reported an error, for that reason.
+	if err := runShutdownHooks(cleanupTimeout); err != nil {
+		log.Error("Cleanup: ", err)
+	}
 	log.Info("Server exiting")
 
 	return nil
 }
 
-// shutdownTimeout is how long Shutdown waits for in-flight requests. It plus
-// whatever cleanup follows has to stay inside the orchestrator's grace period
-// - `docker stop` allows 10s by default before it sends SIGKILL.
-const shutdownTimeout = 5 * time.Second
+// shutdownTimeout is how long Shutdown waits for in-flight requests, and
+// cleanupTimeout how long the BeforeExit callbacks get after it.
+//
+// They are consumed one after the other, so the two together are what has to
+// stay inside the orchestrator's grace period: `docker stop` allows 10s by
+// default before it sends SIGKILL, and 5+3 leaves room for the process to
+// finish returning. Raising either without lowering the other buys nothing -
+// the budget that runs out is the orchestrator's.
+const (
+	shutdownTimeout = 5 * time.Second
+	cleanupTimeout  = 3 * time.Second
+)
 
 // armStopSignals registers for the stop signals and returns the channel they
 // arrive on together with the function that restores the default disposition.
@@ -185,6 +274,66 @@ func armStopSignals() (<-chan os.Signal, func()) {
 	return quit, func() { signal.Stop(quit) }
 }
 
+// startServing binds srv.Addr, hands the listener to srv on its own goroutine,
+// and announces AfterListen.
+//
+// The bind is done here rather than left to ListenAndServe, which binds on the
+// goroutine that serves. That put the failure every deployment actually hits -
+// "address already in use" - on a goroutine nobody was reading, so the banner
+// went on to claim the server was up, and there would be no way to keep
+// AfterListen from announcing a socket that does not exist. A hook there is
+// promised a reachable port; the only way to keep that promise is for the bind
+// to have already happened on this goroutine.
+//
+// AfterListen is announced synchronously. Running it in a goroutine to save the
+// few milliseconds would let it overlap the shutdown: on a fast SIGTERM the
+// cleanup callbacks could finish before the startup ones had.
+//
+// Both ways of failing to start are therefore checked before the announcement:
+// the bind, and - with ssl enabled - the certificate.
+func startServing(srv *http.Server, useTLS bool, pem, key string) error {
+	if useTLS {
+		// Read the certificate before anything is announced. ServeTLS reads
+		// these files itself, but on the serving goroutine - so a bad
+		// certificate used to surface after AfterListen had already promised a
+		// reachable port. Loading it here costs one extra read and moves the
+		// failure onto this goroutine, where run() can return it.
+		//
+		// ServeTLS still does the real work below rather than this handing it a
+		// tls.Listener: that is what sets up HTTP/2 negotiation, and taking it
+		// over here would quietly drop h2 for every TLS deployment.
+		if _, err := tls.LoadX509KeyPair(pem, key); err != nil {
+			return errors.Wrap(err, "tls certificate")
+		}
+	}
+
+	ln, err := net.Listen("tcp", srv.Addr)
+	if err != nil {
+		return errors.Wrap(err, "listen")
+	}
+
+	go func() {
+		// 服务连接
+		var err error
+		if useTLS {
+			err = srv.ServeTLS(ln, pem, key)
+		} else {
+			err = srv.Serve(ln)
+		}
+		if err != nil && !errors.Is(err, http.ErrServerClosed) {
+			// Still fatal, as it was. Neither the bind nor the certificate is
+			// among the errors that reach here any more - both are checked
+			// above, on the caller's goroutine. What is left is a serve that
+			// failed after the port was taken, and carrying on would park the
+			// process on <-quit with nothing serving.
+			log.Fatal("serve: ", err)
+		}
+	}()
+
+	sdk.Runtime.RunPhase(runtime.AfterListen)
+	return nil
+}
+
 // shutdownServer stops srv, giving in-flight requests up to timeout to finish.
 //
 // It returns the error instead of exiting on it. A caller that exits here skips
@@ -196,6 +345,34 @@ func shutdownServer(srv *http.Server, timeout time.Duration) error {
 	return srv.Shutdown(ctx)
 }
 
+// runShutdownHooks runs the BeforeExit callbacks with timeout to share.
+//
+// What the budget bounds is the wait, not the work. When it is gone RunShutdown
+// stops waiting and returns; a callback that never looks at its context carries
+// on until the process exits, and may leave a partial write behind. Go cannot
+// cancel a function that does not check for cancellation, which is why the
+// callbacks are handed a context at all.
+func runShutdownHooks(timeout time.Duration) error {
+	ctx, cancel := context.WithTimeout(context.Background(), timeout)
+	defer cancel()
+	return sdk.Runtime.RunShutdown(ctx)
+}
+
+// buildRouter announces BeforeRouter, builds the engine, and then drains the
+// startup registries.
+//
+// The order is the contract. BeforeRouter is the last point at which a module
+// can still affect how routes are built, so it has to run while there is no
+// engine yet. The before registry runStartupHooks drains is a different moment
+// despite the name: those callbacks run after initRouter has built the engine.
+// Two lines apart, and describing them as equivalent is a mistake this
+// repository has already made once in writing.
+func buildRouter() {
+	sdk.Runtime.RunPhase(runtime.BeforeRouter)
+	initRouter()
+	runStartupHooks()
+}
+
 // runStartupHooks runs the router registries and then the before callbacks.
 //
 // The package-level slice runs first and in its existing order, so a fork that

cmd/api/signal_test.go

@@ -1,6 +1,7 @@
 package api
 
 import (
+	"context"
 	"fmt"
 	"net"
 	"net/http"
@@ -10,6 +11,8 @@ import (
 	"syscall"
 	"testing"
 	"time"
+
+	"github.com/go-admin-team/go-admin-core/v2/sdk"
 )
 
 // The signal path cannot be exercised in-process: delivering a signal to the
@@ -21,13 +24,15 @@ import (
 // this repository's CI has no database (.github/workflows/go.yml runs neither
 // MySQL nor a sqlite-tagged build), and none of what is under test needs one.
 const (
-	childEnv       = "GO_ADMIN_SIGNAL_CHILD"
-	childStuckEnv  = "GO_ADMIN_SIGNAL_CHILD_STUCK"
-	childHangConn  = "GO_ADMIN_SIGNAL_CHILD_HANGCONN"
-	markerReady    = "CHILD-READY"
-	markerSignal   = "CHILD-SIGNAL"
-	markerShutdown = "CHILD-SHUTDOWN-OK"
-	markerExiting  = "CHILD-EXITING"
+	childEnv         = "GO_ADMIN_SIGNAL_CHILD"
+	childStuckEnv    = "GO_ADMIN_SIGNAL_CHILD_STUCK"
+	childHangConn    = "GO_ADMIN_SIGNAL_CHILD_HANGCONN"
+	childSlowCleanup = "GO_ADMIN_SIGNAL_CHILD_SLOWCLEANUP"
+	markerReady      = "CHILD-READY"
+	markerSignal     = "CHILD-SIGNAL"
+	markerShutdown   = "CHILD-SHUTDOWN-OK"
+	markerCleanup    = "CHILD-CLEANUP-RAN"
+	markerExiting    = "CHILD-EXITING"
 )
 
 // TestSignalChild is the child process. It is skipped in a normal run.
@@ -59,6 +64,24 @@ func TestSignalChild(t *testing.T) {
 	}
 	go func() { _ = srv.Serve(ln) }()
 
+	// A BeforeExit callback, registered the way a module would. What the tests
+	// below care about is whether it runs at all - after a Shutdown that
+	// failed, and after its own budget has been spent.
+	cleanupBudget := cleanupTimeout
+	sdk.Runtime.SetShutdown(func(ctx context.Context) {
+		if os.Getenv(childSlowCleanup) == "1" {
+			// Outlasts the budget on purpose, and does not consult ctx -
+			// which is the case the contract is explicit about: what the
+			// context bounds is the wait, not the work.
+			time.Sleep(2 * time.Second)
+		}
+		fmt.Println(markerCleanup)
+		os.Stdout.Sync()
+	})
+	if os.Getenv(childSlowCleanup) == "1" {
+		cleanupBudget = 300 * time.Millisecond
+	}
+
 	// Arm before announcing readiness. Doing it the other way round leaves a
 	// window in which the parent's signal reaches the default handler and
 	// kills the child before any of this runs - which is exactly the failure
@@ -110,6 +133,8 @@ func TestSignalChild(t *testing.T) {
 		timeout = 300 * time.Millisecond
 	}
 
+	sdk.Runtime.BeginShutdown()
+
 	if err := shutdownServer(srv, timeout); err != nil {
 		// Deliberately not fatal, and deliberately not a bare return: the
 		// point is that whatever follows still runs.
@@ -117,6 +142,10 @@ func TestSignalChild(t *testing.T) {
 	} else {
 		fmt.Println(markerShutdown)
 	}
+
+	if err := runShutdownHooks(cleanupBudget); err != nil {
+		fmt.Println("cleanup error:", err)
+	}
 	fmt.Println(markerExiting)
 	os.Stdout.Sync()
 }
@@ -286,7 +315,55 @@ func TestShutdownTimeoutDoesNotStopWhatFollows(t *testing.T) {
 		t.Fatalf("Shutdown did not time out, so this test proves nothing; saw:\n%s",
 			strings.Join(seen, "\n"))
 	}
+	var cleaned bool
+	for _, l := range seen {
+		if strings.Contains(l, markerCleanup) {
+			cleaned = true
+		}
+	}
+	if !cleaned {
+		t.Fatalf("the BeforeExit callback did not run after a failed Shutdown; saw:\n%s",
+			strings.Join(seen, "\n"))
+	}
 	if err := cmd.Wait(); err != nil {
 		t.Fatalf("child exited with %v after a failed Shutdown, want a clean exit", err)
 	}
 }
+
+// A callback that outlasts its budget must not take the process with it, and
+// must not be waited for: RunShutdown reports the deadline and returns, the
+// callback carries on, and the process still exits cleanly. This is the half of
+// the contract that is easy to get backwards - the context bounds the wait, not
+// the work, because Go cannot cancel a function that does not check for it.
+func TestACleanupThatOutlastsItsBudgetIsAbandonedNotAwaited(t *testing.T) {
+	cmd, _, lines := startChild(t, false, childSlowCleanup+"=1")
+	await(t, lines, markerReady, 30*time.Second)
+
+	if err := cmd.Process.Signal(syscall.SIGTERM); err != nil {
+		t.Fatalf("signal: %v", err)
+	}
+	await(t, lines, markerSignal, 10*time.Second)
+
+	// The budget is 300ms and the callback sleeps two seconds. If RunShutdown
+	// waited for it, this marker would not arrive for two seconds; the one
+	// second here is what makes "abandoned, not awaited" the thing asserted.
+	seen := await(t, lines, markerExiting, 1*time.Second)
+
+	var reported bool
+	for _, l := range seen {
+		if strings.Contains(l, "cleanup error:") {
+			reported = true
+		}
+		if strings.Contains(l, markerCleanup) {
+			t.Fatalf("the slow callback finished before the process moved on, so nothing was abandoned; saw:\n%s",
+				strings.Join(seen, "\n"))
+		}
+	}
+	if !reported {
+		t.Fatalf("RunShutdown returned no error for a callback that outlasted the budget; saw:\n%s",
+			strings.Join(seen, "\n"))
+	}
+	if err := cmd.Wait(); err != nil {
+		t.Fatalf("child exited with %v, want a clean exit despite the abandoned callback", err)
+	}
+}

common/health/health.go

@@ -0,0 +1,157 @@
+// Package health answers whether this process should be sent traffic.
+//
+// The two questions an orchestrator asks are not the same one, and go-admin
+// answers them at two endpoints:
+//
+//   - /health is liveness: is the process there at all. It stays a bare 200,
+//     because the honest answer to "should I restart you" is almost always no.
+//     Restarting a process because its database is unreachable turns one
+//     outage into a crash loop that also loses the connection pool, the cache
+//     and every in-flight request.
+//   - /ready is readiness: should this instance receive requests now. It fails
+//     while the dependencies are unreachable, and - the part that only exists
+//     because of the life-cycle phases - it fails as soon as shutdown begins,
+//     before the server stops accepting, so a load balancer has a chance to
+//     take the instance out before connections are cut.
+package health
+
+import (
+	"context"
+	"crypto/rand"
+	"encoding/hex"
+	"errors"
+	"fmt"
+	"sync/atomic"
+	"time"
+
+	"github.com/go-admin-team/go-admin-core/v2/sdk"
+)
+
+// draining is set when the process starts shutting down.
+//
+// It is kept here rather than read back from core: BeginShutdown sets a flag on
+// the Application, but nothing exports it, and one host wanting to know is not
+// yet a reason to widen that interface.
+var draining atomic.Bool
+
+// BeginDraining records that shutdown has started, so readiness fails from now
+// on. It is called with BeginShutdown, before anything is taken apart.
+func BeginDraining() { draining.Store(true) }
+
+// Draining reports whether shutdown has begun.
+func Draining() bool { return draining.Load() }
+
+// Check is one dependency and what asking it produced.
+type Check struct {
+	Name string `json:"name"`
+	OK   bool   `json:"ok"`
+	Err  string `json:"error,omitempty"`
+}
+
+// Ready asks every dependency this process cannot serve a request without.
+//
+// The queue is deliberately absent. Nothing on AdapterQueue answers "are you
+// reachable" without publishing something, the memory backend cannot fail, and
+// a queue that is down degrades logging rather than stopping requests - which
+// is a reason to alert, not a reason to leave the load balancer pool.
+func Ready(ctx context.Context) []Check {
+	return []Check{
+		safely("database", func() error { return pingDB(ctx) }),
+		safely("cache", probeCache),
+	}
+}
+
+// safely turns a panic into a failed check.
+//
+// Not defensive habit: the accessors hand back wrappers, not the resources.
+// sdk.Runtime.GetCacheAdapter builds a runtime.Cache around whatever is
+// configured and returns it even when nothing is - so the value is not nil, the
+// cache inside it is, and the first call dereferences it. A nil check cannot
+// see that, and the same is true of GetQueueAdapter.
+//
+// Whatever the reason, a probe is the last thing that should be able to take
+// the process down: the caller is asking whether this instance is well, and
+// killing it to answer is the wrong reply.
+func safely(name string, fn func() error) (c Check) {
+	c = Check{Name: name}
+	defer func() {
+		if r := recover(); r != nil {
+			c.OK, c.Err = false, fmt.Sprintf("the check panicked: %v", r)
+		}
+	}()
+	if err := fn(); err != nil {
+		c.Err = err.Error()
+		return c
+	}
+	c.OK = true
+	return c
+}
+
+// Healthy reports whether every check passed.
+func Healthy(checks []Check) bool {
+	for _, c := range checks {
+		if !c.OK {
+			return false
+		}
+	}
+	return true
+}
+
+func pingDB(ctx context.Context) error {
+	db := sdk.Runtime.GetDb()
+	if db == nil {
+		return errors.New("no database configured")
+	}
+	sqlDB, err := db.DB()
+	if err != nil {
+		return err
+	}
+	return sqlDB.PingContext(ctx)
+}
+
+// cacheProbePrefix names the probe's keys. The key itself is per probe, not
+// fixed: two /ready requests arriving together - or two instances sharing one
+// redis, which is the normal deployment - would otherwise overwrite each
+// other's value between the write and the read and each conclude the cache was
+// broken. A readiness probe that reports false negatives under load takes
+// healthy instances out of the pool, which is worse than not probing.
+const cacheProbePrefix = "go-admin:health:"
+
+// cacheProbeTTL is short because these keys are write-once and never read
+// again by anyone else; it only has to outlive the read that follows.
+const cacheProbeTTL = 30
+
+func probeCache() error {
+	adapter := sdk.Runtime.GetCacheAdapter()
+	if adapter == nil {
+		return errors.New("no cache configured")
+	}
+
+	suffix := make([]byte, 8)
+	if _, err := rand.Read(suffix); err != nil {
+		return fmt.Errorf("could not build a probe key: %w", err)
+	}
+	key := cacheProbePrefix + hex.EncodeToString(suffix)
+
+	// Written and read back rather than only read: a cache that answers "miss"
+	// for every key - a client pointed at the wrong server - is
+	// indistinguishable from a healthy one on a read alone.
+	want := time.Now().Format(time.RFC3339Nano)
+	if err := adapter.Set(key, want, cacheProbeTTL); err != nil {
+		return err
+	}
+	// Best effort, and its error is deliberately dropped: the verdict is
+	// already decided by the read below, and a cache that cannot delete a key
+	// it just wrote is not a reason to refuse traffic. The TTL is the real
+	// cleanup.
+	defer func() { _ = adapter.Del(key) }()
+
+	got, err := adapter.Get(key)
+	if err != nil {
+		return err
+	}
+	if got != want {
+		return errors.New("the cache returned a different value than was written")
+	}
+	return nil
+}

common/health/health_test.go

@@ -0,0 +1,241 @@
+package health
+
+import (
+	"context"
+	"errors"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/go-admin-team/go-admin-core/v2/sdk"
+	"github.com/go-admin-team/go-admin-core/v2/sdk/runtime"
+	corestorage "github.com/go-admin-team/go-admin-core/v2/storage"
+)
+
+func freshRuntime(t *testing.T) {
+	t.Helper()
+	previous := sdk.Runtime
+	t.Cleanup(func() { sdk.Runtime = previous })
+	sdk.Runtime = runtime.NewConfig()
+}
+
+// fakeCache answers whatever the test needs it to.
+type fakeCache struct {
+	mu      sync.Mutex
+	setErr  error
+	getErr  error
+	getBack string // returned instead of what was written, when non-empty
+	stored  map[string]string
+
+	// oneSlot makes the cache keep a single value however many keys are
+	// written, which is what a shared probe key turns any cache into.
+	oneSlot bool
+	slot    string
+
+	// setBarrier, when set, holds every writer until all of them have written.
+	// Without it the probes are short enough that the scheduler usually runs
+	// them one after another, and a shared key survives by luck rather than by
+	// design - which would leave the test below asserting nothing.
+	setBarrier *barrier
+}
+
+// barrier releases every waiter once n of them have arrived.
+type barrier struct {
+	n   int
+	mu  sync.Mutex
+	got int
+	ch  chan struct{}
+}
+
+func newBarrier(n int) *barrier { return &barrier{n: n, ch: make(chan struct{})} }
+
+func (b *barrier) wait() {
+	b.mu.Lock()
+	b.got++
+	if b.got == b.n {
+		close(b.ch)
+	}
+	b.mu.Unlock()
+	<-b.ch
+}
+
+func (c *fakeCache) String() string { return "fake" }
+
+func (c *fakeCache) Set(key string, val interface{}, _ int) error {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	if c.setErr != nil {
+		return c.setErr
+	}
+	v, _ := val.(string)
+	if c.oneSlot {
+		c.slot = v
+		return nil
+	}
+	if c.stored == nil {
+		c.stored = map[string]string{}
+	}
+	c.stored[key] = v
+	c.mu.Unlock()
+	if c.setBarrier != nil {
+		// Outside the lock on purpose: waiting while holding it would deadlock
+		// every other writer before the barrier could fill.
+		c.setBarrier.wait()
+	}
+	c.mu.Lock()
+	return nil
+}
+
+func (c *fakeCache) Get(key string) (string, error) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	if c.getErr != nil {
+		return "", c.getErr
+	}
+	if c.getBack != "" {
+		return c.getBack, nil
+	}
+	if c.oneSlot {
+		return c.slot, nil
+	}
+	return c.stored[key], nil
+}
+
+func (c *fakeCache) Del(key string) error {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	delete(c.stored, key)
+	return nil
+}
+
+func (c *fakeCache) HashGet(_, _ string) (string, error) { return "", nil }
+func (c *fakeCache) HashDel(_, _ string) error           { return nil }
+func (c *fakeCache) Increase(string) error               { return nil }
+func (c *fakeCache) Decrease(string) error               { return nil }
+func (c *fakeCache) Expire(string, time.Duration) error  { return nil }
+
+var _ corestorage.AdapterCache = (*fakeCache)(nil)
+
+func named(checks []Check, name string) Check {
+	for _, c := range checks {
+		if c.Name == name {
+			return c
+		}
+	}
+	return Check{Name: name, Err: "check not reported at all"}
+}
+
+// A cache that accepts writes and answers every read with a different value is
+// the failure this probe exists for - a client pointed at the wrong server, or
+// one that silently drops everything. A read alone cannot tell that apart from
+// a healthy cache with a cold key, which is why the probe writes first.
+func TestCacheProbeFailsWhenTheValueDoesNotComeBack(t *testing.T) {
+	freshRuntime(t)
+	sdk.Runtime.SetCacheAdapter(&fakeCache{getBack: "something else"})
+
+	got := named(Ready(context.Background()), "cache")
+	if got.OK {
+		t.Error("the cache check passed although the value written was not the value read back")
+	}
+	if got.Err == "" {
+		t.Error("the failing check reported no reason")
+	}
+}
+
+func TestCacheProbePassesWhenTheValueComesBack(t *testing.T) {
+	freshRuntime(t)
+	sdk.Runtime.SetCacheAdapter(&fakeCache{})
+
+	if got := named(Ready(context.Background()), "cache"); !got.OK {
+		t.Errorf("the cache check failed for a cache that works: %s", got.Err)
+	}
+}
+
+func TestCacheProbeReportsAWriteFailure(t *testing.T) {
+	freshRuntime(t)
+	sdk.Runtime.SetCacheAdapter(&fakeCache{setErr: errors.New("connection refused")})
+
+	got := named(Ready(context.Background()), "cache")
+	if got.OK {
+		t.Error("the cache check passed although the write failed")
+	}
+}
+
+// Nothing configured is the case that used to take the process down rather
+// than answer. GetCacheAdapter builds a wrapper around whatever is configured
+// and returns it even when nothing is, so the value is not nil, the cache
+// inside it is, and Set dereferences it - a probe that panics is the worst
+// possible answer to "are you well".
+//
+// Every check has to be reported, passing or not. A probe that omits what it
+// could not reach reads as a shorter list of healthy things.
+func TestEveryDependencyIsReportedEvenWithNothingConfigured(t *testing.T) {
+	freshRuntime(t)
+
+	checks := Ready(context.Background())
+	for _, name := range []string{"database", "cache"} {
+		c := named(checks, name)
+		if c.Err == "check not reported at all" {
+			t.Errorf("%s was not reported", name)
+		}
+		if c.OK {
+			t.Errorf("%s passed with nothing configured", name)
+		}
+	}
+	if Healthy(checks) {
+		t.Error("Healthy said yes for a process with no database and no cache")
+	}
+}
+
+// Draining is what makes the shutdown graceful from the outside: it has to be
+// observable before the server stops accepting, or the load balancer learns
+// about the shutdown by having its connections cut.
+func TestDrainingIsObservableOnceItBegins(t *testing.T) {
+	previous := draining.Load()
+	t.Cleanup(func() { draining.Store(previous) })
+
+	draining.Store(false)
+	if Draining() {
+		t.Fatal("Draining reported true before shutdown began")
+	}
+	BeginDraining()
+	if !Draining() {
+		t.Error("Draining still reported false after BeginDraining")
+	}
+}
+
+// Two probes at once must both pass. With one fixed key they overwrite each
+// other's value between the write and the read, and a readiness probe that
+// reports false negatives under load takes healthy instances out of the pool -
+// which is worse than not probing at all.
+func TestConcurrentProbesDoNotOverwriteEachOther(t *testing.T) {
+	freshRuntime(t)
+	const probes = 16
+	sdk.Runtime.SetCacheAdapter(&fakeCache{setBarrier: newBarrier(probes)})
+
+	var wg sync.WaitGroup
+	failures := make(chan string, probes)
+	for i := 0; i < probes; i++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			if c := named(Ready(context.Background()), "cache"); !c.OK {
+				failures <- c.Err
+			}
+		}()
+	}
+	wg.Wait()
+	close(failures)
+
+	var n int
+	var first string
+	for err := range failures {
+		if n == 0 {
+			first = err
+		}
+		n++
+	}
+	if n > 0 {
+		t.Errorf("%d of %d concurrent probes called a healthy cache broken; first: %s", n, probes, first)
+	}
+}

common/middleware/settings.go

@@ -34,6 +34,7 @@ var CasbinExclude = []UrlInfo{
 	{Url: "/api/v1/user/pwd", Method: "PUT"},
 	{Url: "/api/v1/metrics", Method: "GET"},
 	{Url: "/api/v1/health", Method: "GET"},
+	{Url: "/api/v1/ready", Method: "GET"},
 	{Url: "/", Method: "GET"},
 	{Url: "/api/v1/server-monitor", Method: "GET"},
 	{Url: "/api/v1/public/uploadFile", Method: "POST"},

common/storage/initialize.go

@@ -9,10 +9,11 @@ package storage
 
 import (
 	"log"
+	"sync"
 
+	"github.com/go-admin-team/go-admin-core/v2/captcha"
 	"github.com/go-admin-team/go-admin-core/v2/sdk"
 	"github.com/go-admin-team/go-admin-core/v2/sdk/config"
-	"github.com/go-admin-team/go-admin-core/v2/captcha"
 )
 
 // Setup 配置storage组件
@@ -34,17 +35,75 @@ func setupCaptcha() {
 	captcha.SetStore(captcha.NewCacheStore(sdk.Runtime.GetCacheAdapter(), 600))
 }
 
+var (
+	queueMu sync.Mutex
+	// installed is the adapter setupQueue built, kept so the next reload can
+	// shut it down, and counted so a consumer can tell one from the next.
+	installed    interface{ Shutdown() }
+	installedGen uint64
+)
+
+// QueueGeneration reports how many times this package has installed a queue
+// adapter. It changes every time setupQueue builds a new one, which is on
+// every configuration reload, and stays 0 for as long as the configuration has
+// no queue section at all - in which case nothing is installed and callers are
+// working with the runtime's own fallback queue.
+//
+// It exists because there is no way to ask for the adapter's identity from the
+// outside. sdk.Runtime.GetQueueAdapter and GetQueuePrefix build a fresh
+// runtime.Queue wrapper on every call, so comparing what two calls return
+// compares two wrappers and never matches, however many times the underlying
+// adapter has been replaced. This package creates the adapter, so this is the
+// only place that knows. A counter rather than the adapter itself keeps the
+// comparison on a uint64: an adapter type that is not comparable would panic
+// an `==` between two interface values.
+func QueueGeneration() uint64 {
+	queueMu.Lock()
+	defer queueMu.Unlock()
+	return installedGen
+}
+
 func setupQueue() {
 	if config.QueueConfig.Empty() {
 		return
 	}
-	if q := sdk.Runtime.GetQueueAdapter(); q != nil {
-		q.Shutdown()
-	}
+
+	queueMu.Lock()
+	defer queueMu.Unlock()
+
 	queueAdapter, err := config.QueueConfig.Setup()
 	if err != nil {
 		log.Fatalf("queue setup error, %s\n", err.Error())
 	}
+
+	previous := installed
 	sdk.Runtime.SetQueueAdapter(queueAdapter)
-	go queueAdapter.Run()
+	installed = queueAdapter
+	installedGen++
+
+	// The previous adapter goes down after the new one is installed, not
+	// before. Shutdown waits for its consumers to deliver what it still holds,
+	// and for that whole wait the runtime would otherwise be handing producers
+	// a queue that has stopped accepting: every Append in the window comes back
+	// ErrQueueClosed, and both call sites in common/middleware log it. Swapping
+	// first leaves no such window - a producer gets the new queue or the old
+	// one, and both work.
+	//
+	// Only an adapter this package installed. GetQueueAdapter never returns
+	// nil - with nothing configured the runtime falls back to its own memory
+	// queue and wraps that - so the `if q := GetQueueAdapter(); q != nil` this
+	// replaces was always true, and shut down the fallback queue on the very
+	// first start, before anything had used it.
+	if previous != nil {
+		previous.Shutdown()
+	}
+
+	// Deliberately not started here. Run has to come after the consumers have
+	// registered: the contract implementations refuse a registration once the
+	// queue is running (storage.ErrQueueAlreadyStarted), and the legacy
+	// adapter this repository still goes through swallows that error rather
+	// than reporting it - its own comment says the interface gives it no way
+	// to tell the caller. Starting here and registering afterwards is
+	// therefore a race that loses consumers in silence. Whoever registers is
+	// the one that starts it.
 }

common/storage/queue_redis_test.go

@@ -0,0 +1,130 @@
+package storage
+
+import (
+	"errors"
+	"os"
+	"testing"
+	"time"
+
+	"github.com/go-admin-team/go-admin-core/v2/sdk/config"
+	corestorage "github.com/go-admin-team/go-admin-core/v2/storage"
+	"github.com/go-admin-team/go-admin-core/v2/storage/queue"
+)
+
+// redisAddrEnv points these tests at a server. They are skipped without it, so
+// a developer with no redis running still gets a green run - and CI sets it,
+// which is the point: the ordering rule they cover is invisible on the memory
+// backend, and memory is the default. A suite that only ever exercised the
+// default would report success for a queue that silently drops every consumer.
+const redisAddrEnv = "GO_ADMIN_TEST_REDIS_ADDR"
+
+func redisAddr(t *testing.T) string {
+	t.Helper()
+	addr := os.Getenv(redisAddrEnv)
+	if addr != "" {
+		return addr
+	}
+	// Skipping locally is the point; skipping in CI is the failure this whole
+	// file exists to prevent. A workflow that renamed the variable, or dropped
+	// the service, would otherwise go green while these two tests quietly did
+	// nothing - which is the same shape as the defect they cover.
+	if os.Getenv("CI") != "" {
+		t.Fatalf("%s is not set while CI is: the redis-backed queue tests must not skip here", redisAddrEnv)
+	}
+	t.Skipf("%s is not set; skipping the redis-backed queue tests", redisAddrEnv)
+	return ""
+}
+
+// newRedisQueue builds the queue the same way setupQueue does - through
+// config.QueueConfig.Setup - so that what is under test is the adapter this
+// repository actually gets, LegacyQueueAdapter and all, rather than a redis
+// client wired up by the test.
+func newRedisQueue(t *testing.T, prefix string) corestorage.AdapterQueue {
+	t.Helper()
+	previous := config.QueueConfig
+	t.Cleanup(func() { config.QueueConfig = previous })
+
+	config.QueueConfig = &config.Queue{
+		Redis: &config.RedisQueue{
+			RedisOptions: config.RedisOptions{Addr: redisAddr(t)},
+			Group:        prefix,
+			KeyPrefix:    prefix,
+		},
+	}
+	q, err := config.QueueConfig.Setup()
+	if err != nil {
+		t.Fatalf("queue setup: %v", err)
+	}
+	t.Cleanup(q.Shutdown)
+	return q
+}
+
+func message(t *testing.T, stream string) corestorage.Messager {
+	t.Helper()
+	m := &queue.Message{}
+	m.SetStream(stream)
+	m.SetValues(map[string]interface{}{"hello": "world"})
+	return m
+}
+
+// Registered first, then started: the consumer gets the message. This is the
+// order setupQueue and attachQueueConsumers now produce between them.
+func TestRedisQueueDeliversToAConsumerRegisteredBeforeTheStart(t *testing.T) {
+	stream := "t-ordered"
+	q := newRedisQueue(t, "gotest-ordered")
+
+	got := make(chan struct{}, 1)
+	q.Register(stream, func(corestorage.Messager) error {
+		select {
+		case got <- struct{}{}:
+		default:
+		}
+		return nil
+	})
+	go q.Run()
+
+	// Give Start a moment to reach its read loop before publishing.
+	time.Sleep(500 * time.Millisecond)
+	if err := q.Append(message(t, stream)); err != nil {
+		t.Fatalf("append: %v", err)
+	}
+
+	select {
+	case <-got:
+	case <-time.After(15 * time.Second):
+		t.Fatal("the consumer never received the message")
+	}
+}
+
+// Started first, then registered: the registration is refused and every
+// publish afterwards fails.
+//
+// Subscribe answers ErrQueueAlreadyStarted, and LegacyQueueAdapter.Register
+// returns nothing, so the caller cannot know - that part is silent. What is not
+// silent is the consequence: no consumer group was created, so Publish refuses
+// the topic with ErrNoHandler on every single request, and go-admin's call
+// sites log that at error level while the login and operation log rows are
+// never written.
+//
+// This is the test the memory backend cannot provide. queue.Memory's Register
+// starts another consumer goroutine whatever the state, so the same code passes
+// there - which is how the defect survived, memory being the default.
+func TestRedisQueueRefusesAConsumerRegisteredAfterTheStart(t *testing.T) {
+	stream := "t-late"
+	q := newRedisQueue(t, "gotest-late")
+
+	go q.Run()
+	time.Sleep(500 * time.Millisecond)
+
+	q.Register(stream, func(corestorage.Messager) error { return nil })
+
+	err := q.Append(message(t, stream))
+	if err == nil {
+		t.Fatal("a message was accepted for a topic whose registration came after Start; " +
+			"if the backend now accepts late registration, the ordering rule in setupQueue can be revisited")
+	}
+	if !errors.Is(err, corestorage.ErrNoHandler) {
+		t.Fatalf("append failed with %v, want %v - the test is meant to pin the "+
+			"missing-consumer path, not any error at all", err, corestorage.ErrNoHandler)
+	}
+}

common/storage/queue_swap_test.go

@@ -0,0 +1,161 @@
+package storage
+
+import (
+	"errors"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/go-admin-team/go-admin-core/v2/sdk"
+	"github.com/go-admin-team/go-admin-core/v2/sdk/config"
+	"github.com/go-admin-team/go-admin-core/v2/sdk/runtime"
+	corestorage "github.com/go-admin-team/go-admin-core/v2/storage"
+	"github.com/go-admin-team/go-admin-core/v2/storage/queue"
+)
+
+// sampleSize is how many publishes have to land inside the reload before the
+// measurement is taken. Waiting on the count rather than on wall clock keeps
+// the window the test covers the same on a loaded runner as on an idle one.
+const sampleSize = 200
+
+func swapMsg() corestorage.Messager {
+	m := new(queue.Message)
+	m.SetStream("t")
+	m.SetValues(map[string]interface{}{"a": "b"})
+	return m
+}
+
+// A reload must never leave producers holding a queue that has stopped
+// accepting.
+//
+// Shutdown waits for its consumers to deliver what the queue still holds. Taking
+// the old adapter down before installing the new one meant the runtime pointed
+// at a closed queue for that entire wait: every Append in the window came back
+// ErrQueueClosed, and both call sites in common/middleware log it at error
+// level. Installing first leaves no window - a producer gets the new queue or
+// the old one, and both accept.
+//
+// The difference is only visible during that wait, which is why the test holds
+// a consumer rather than checking the state after Setup has returned: by then
+// the two orders look identical.
+//
+// One refusal survives the fix and is not something this ordering can reach.
+// GetQueuePrefix hands back a wrapper that captured the adapter, so a producer
+// that fetched before the swap and appends after Shutdown has begun is still
+// holding the old one. That window is one call wide and closing it means
+// resolving the adapter inside Append, which is core's to change. What the
+// ordering removes is the sustained window: every producer that fetches during
+// the wait. The test publishes from a single goroutine, so at most one of its
+// calls can straddle the swap - which is what makes "more than one" the line
+// between the two orders rather than a tolerance.
+func TestAReloadNeverPointsProducersAtAClosedQueue(t *testing.T) {
+	prevQ, prevC := config.QueueConfig, config.CacheConfig
+	prevRuntime := sdk.Runtime
+	prevInstalled, prevGen := installed, installedGen
+	t.Cleanup(func() {
+		config.QueueConfig, config.CacheConfig = prevQ, prevC
+		sdk.Runtime = prevRuntime
+		queueMu.Lock()
+		installed, installedGen = prevInstalled, prevGen
+		queueMu.Unlock()
+	})
+	sdk.Runtime = runtime.NewConfig()
+	config.CacheConfig = &config.Cache{Memory: struct{}{}}
+	// Sized so the buffer cannot fill while the consumer is held: a full queue
+	// returns an error of its own, and this test needs every error other than
+	// ErrQueueClosed to mean something it does not model has happened.
+	config.QueueConfig = &config.Queue{Memory: &config.QueueMemory{PoolSize: 4096}}
+
+	Setup()
+
+	// A consumer that will not finish until this test lets it, so the reload's
+	// Shutdown has something to wait for.
+	release := make(chan struct{})
+	consuming := make(chan struct{})
+	var picked sync.Once
+	first := sdk.Runtime.GetQueuePrefix("")
+	first.Register("t", func(corestorage.Messager) error {
+		picked.Do(func() { close(consuming) })
+		<-release
+		return nil
+	})
+	go first.Run()
+	for i := 0; i < 4; i++ {
+		if err := first.Append(swapMsg()); err != nil {
+			t.Fatalf("seed append %d: %v", i, err)
+		}
+	}
+	select {
+	case <-consuming:
+	case <-time.After(10 * time.Second):
+		t.Fatal("the consumer never picked a message up, so the reload has nothing to wait for")
+	}
+
+	reloaded := make(chan struct{})
+	go func() { Setup(); close(reloaded) }()
+
+	// Publish continuously while the reload is in progress.
+	var refused atomic.Int64
+	var attempts atomic.Int64
+	unexpected := make(chan error, 1)
+	stop := make(chan struct{})
+	// publishing is closed by the producer on its way out. The test joins on it
+	// before returning: t.Cleanup restores sdk.Runtime, and a producer still in
+	// flight would be reading the variable that restore writes.
+	publishing := make(chan struct{})
+	go func() {
+		defer close(publishing)
+		for {
+			select {
+			case <-stop:
+				return
+			default:
+			}
+			attempts.Add(1)
+			err := sdk.Runtime.GetQueuePrefix("").Append(swapMsg())
+			switch {
+			case err == nil:
+			case errors.Is(err, corestorage.ErrQueueClosed):
+				refused.Add(1)
+			default:
+				// Kept rather than counted: an Append refused for some other
+				// reason would otherwise leave refused at zero and the test
+				// green while nothing was reaching a queue at all.
+				select {
+				case unexpected <- err:
+				default:
+				}
+			}
+			time.Sleep(time.Millisecond)
+		}
+	}()
+
+	deadline := time.After(30 * time.Second)
+	for attempts.Load() < sampleSize {
+		select {
+		case <-deadline:
+			t.Fatalf("only %d publishes landed inside the reload; the window was never sampled", attempts.Load())
+		case <-time.After(time.Millisecond):
+		}
+	}
+	close(release)
+
+	select {
+	case <-reloaded:
+	case <-time.After(30 * time.Second):
+		t.Fatal("the reload never finished")
+	}
+	close(stop)
+	<-publishing
+
+	select {
+	case err := <-unexpected:
+		t.Fatalf("a publish failed for a reason this test does not model: %v", err)
+	default:
+	}
+	if n := refused.Load(); n > 1 {
+		t.Errorf("%d of %d publishes during the reload were refused: producers were pointed at the closed queue",
+			n, attempts.Load())
+	}
+}

common/storage/setup_gen_test.go

@@ -0,0 +1,53 @@
+package storage
+
+import (
+	"testing"
+
+	"github.com/go-admin-team/go-admin-core/v2/sdk"
+	"github.com/go-admin-team/go-admin-core/v2/sdk/config"
+	"github.com/go-admin-team/go-admin-core/v2/sdk/runtime"
+)
+
+// Issue #892: a configuration reload replaces the queue adapter and the
+// consumers registered against the previous one are attached to a queue nobody
+// publishes to any more.
+//
+// The fix has two halves. attachQueueConsumers gives a new queue its own
+// consumers and the same queue none, which cmd/api covers against a queue the
+// test controls. This is the other half: that a reload actually produces a new
+// queue for it to notice. Setup is what config re-runs on every change, so
+// calling it twice is what a reload does to this package.
+func TestSetupBumpsTheQueueGenerationOnEveryReload(t *testing.T) {
+	// Setup writes the process-wide sdk.Runtime - the cache and queue adapters -
+	// and this package's own record of what it installed. Restoring all of it
+	// keeps the test from deciding what a later test in this binary sees,
+	// which is the same isolation cmd/api's freshRuntime provides.
+	prevQ, prevC := config.QueueConfig, config.CacheConfig
+	prevRuntime := sdk.Runtime
+	prevInstalled, prevGen := installed, installedGen
+	t.Cleanup(func() {
+		config.QueueConfig, config.CacheConfig = prevQ, prevC
+		sdk.Runtime = prevRuntime
+		queueMu.Lock()
+		installed, installedGen = prevInstalled, prevGen
+		queueMu.Unlock()
+	})
+	sdk.Runtime = runtime.NewConfig()
+
+	config.CacheConfig = &config.Cache{Memory: struct{}{}}
+	config.QueueConfig = &config.Queue{Memory: &config.QueueMemory{PoolSize: 10}}
+
+	before := QueueGeneration()
+	Setup()
+	first := QueueGeneration()
+	Setup()
+	second := QueueGeneration()
+
+	t.Logf("before=%d first=%d second=%d", before, first, second)
+	if first == before {
+		t.Fatal("the first Setup did not install a queue")
+	}
+	if second == first {
+		t.Fatal("a second Setup - which is what a configuration reload does - did not install a new one")
+	}
+}

docs/contract.md

@@ -377,7 +377,7 @@ if res.RowsAffected == 0 { return ErrAlreadyPaid }   // 别人先改了
 | 应用间调用 | 零定义。A 应用要调 B 应用只能直接 import 对方的包，循环依赖就回来了 |
 | 领域事件 / EventBus | 无 |
 | 缓存的租户隔离 | `service.Service` 有 `Cache` 字段，**是否按租户隔离未验证**。当作没隔离来写 |
-| 异步任务 | 有队列，但热更新后消费者会丢（issue #892） |
+| 生命周期钩子之外的时点 | 只有下面那四个。没有「路由装好之后、开始监听之前」这一档 |
 
 这几条留给后续批次，按真实需求补——现在凭空设计只会设计错。
 如果你的应用卡在这里，在 issue 里说一声，那正是我们要的输入。
@@ -672,6 +672,91 @@ if sdk.Runtime.AppRoutersSealed() { /* RunAppRouters 已经跑过了 */ }
 
 ---
 
+## 生命周期挂载点
+
+除了注册路由和迁移，应用还可以把工作挂在进程生命的四个时点上，不必等宿主按名字来调自己。
+契约本身在 core，见
+[go-admin-core `docs/contract.md`](https://github.com/go-admin-team/go-admin-core/blob/main/docs/contract.md)
+的「Life-cycle phases」一节。这里只写**在本仓里它们分别落在哪一行**。
+
+| 阶段 | 在 `cmd/api/server.go` 的位置 | 此时可用 |
+|---|---|---|
+| `AfterResource` | `bootstrap.SetupConfig` 跑完 `database.Setup` / `storage.Setup` 之后 | 配置、库、缓存、队列、casbin |
+| `BeforeRouter` | `initRouter()` **之前** | 以上，加引擎尚未构建这一事实 |
+| `AfterListen` | `startServing()` 里，`net.Listen` 返回之后 | 全部，端口**已绑定**、连接进得来 |
+| `BeforeExit` | `srv.Shutdown` 返回之后（无论它是否报错） | 全部，正在被拆掉 |
+
+`AfterListen` 承诺的是**端口已绑定**，不是「`Serve` 已经在 accept 循环里」——
+`srv.Serve` 在另一个 goroutine 上。这个区别是真实的：绑定成功之后内核就会把连接排进
+backlog，所以钩子里去连自己的端口不会被拒;但此刻 `Serve` 可能还没跑到第一次 `Accept`。
+绑定失败则**根本不会有这个阶段**：`net.Listen` 的错误直接从 `run()` 返回，
+横幅不打印，进程非零退出。
+
+```go
+sdk.Runtime.SetPhase(runtime.AfterResource, func() { /* ... */ })
+sdk.Runtime.SetShutdown(func(ctx context.Context) { /* ... */ })
+```
+
+### `BeforeRouter` 不等于 `before` 注册表
+
+**这两个不是同一个时点，文档里别混着写。** `SetBefore` 的回调由
+`runStartupHooks()` 执行，而那是在 `initRouter()` **之后**——引擎已经建好了。
+`BeforeRouter` 在它之前。
+
+顺带：`BeforeRouter` 是「硬约束：注册要赶在启动钩子之前」那一节所说的合法注册窗口之一。
+它早于 `runStartupHooks()`，所以在这里调 `sdk.Runtime.SetAppRouters` 仍然来得及。
+
+### `AfterResource` 会跑很多次，回调必须扛得住
+
+它在**每次配置热更新之后**都会再跑一遍，因为热更新会重建它所命名的那些资源。
+所以这里的回调要求是**「对同一个资源幂等」，不是「第二次什么都不做」**。
+
+本仓自己的队列消费者就是这条规则的样板，也是它存在的理由
+（`cmd/api/server.go` 的 `attachQueueConsumers`）：
+
+- 热更新重建了队列适配器，挂在旧适配器上的消费者连着一个**再没人往里发消息**的队列，
+  登录日志和操作日志就此停写且不出声。所以新适配器**必须**重新注册。
+- 但同一个适配器不能注册两次，否则每条消息有两个消费者，每行日志写两遍。
+
+**身份不能从访问器取。** `sdk.Runtime.GetQueueAdapter()` 与 `GetQueuePrefix()`
+每次调用都新造一个 `runtime.Queue` 包装，比较两次返回等于比较两个包装，
+**底层适配器换过多少次都不相等**。要在**创建资源的地方**记身份——
+本仓是 `common/storage.QueueGeneration()`。
+
+### 注册消费者要赶在队列启动之前
+
+走哪条实现，取决于配置里有没有 `redis:` 段（`config.QueueConfig.Setup()`）：
+
+| 配置 | 实际类型 | 启动后还能注册吗 |
+|---|---|---|
+| `queue: memory:` | `queue.NewMemory` | **能**。它的 `Register` 每次起一个消费 goroutine，不看是否已 `Run` |
+| `queue: redis:` | `storage.LegacyQueueAdapter` 包着新契约实现 | **不能**。`Register` 内部调 `Subscribe`，启动后返回 `storage.ErrQueueAlreadyStarted` |
+
+而 `LegacyQueueAdapter.Register` **没有返回值**——它只能把这个错误写进 slog，
+core 里那行注释自己写着「The interface has no way to report this to the caller」。
+**静默的是注册这一步，不是之后。** 没有建立消费组，redis 会用
+`storage.ErrNoHandler` 拒绝**之后的每一次投递**，而本仓两个调用点
+（`common/middleware/logger.go`、`common/middleware/handler/auth.go`）
+都把它记为 error——于是日志行一条都不落库，同时每个请求刷一条错误日志。
+
+所以顺序是硬的：**先 `Register` 完，再由注册方 `Run()`。**
+`common/storage` 的 `setupQueue` 有意不启动队列。
+
+默认配置选的是 memory 后端,它不在乎顺序——**这个缺陷在默认部署里看不见,
+只在配了 redis 的部署上发作**,而丢掉的正是登录日志、操作日志和 api 检查。
+
+### `BeforeExit` 反序执行，预算约束的是等待
+
+清理按**注册的逆序**执行。`SetShutdown` 拿到宿主剩余的预算，
+但**它约束的是等待，不是工作**：预算用尽时 `RunShutdown` 停止等待并返回，
+而不检查 context 的回调会一直跑到进程退出。Go 没法取消一个不检查取消的函数。
+
+本仓的样板是 cron（`app/jobs/jobbase.go` 的 `startCrontab`）：
+`cron.Stop()` 返回一个在**已经在跑的任务结束时**关闭的 context，
+钩子在它和预算之间二选一。
+
+---
+
 ## 安全边界：装一个应用等于信任它
 
 **这一层划不出安全边界，本文不假装划得出。**

go.mod

@@ -11,7 +11,7 @@ require (
 	github.com/casbin/casbin/v3 v3.8.1
 	github.com/gin-gonic/gin v1.12.0
 	github.com/glebarez/sqlite v1.11.0
-	github.com/go-admin-team/go-admin-core/v2 v2.5.0
+	github.com/go-admin-team/go-admin-core/v2 v2.7.0
 	github.com/google/uuid v1.6.0
 	github.com/huaweicloud/huaweicloud-sdk-go-obs v3.26.6+incompatible
 	github.com/mssola/user_agent v0.6.0

go.sum

@@ -147,6 +147,10 @@ github.com/glebarez/sqlite v1.11.0 h1:wSG0irqzP6VurnMEpFGer5Li19RpIRi2qvQz++w0GM
 github.com/glebarez/sqlite v1.11.0/go.mod h1:h8/o8j5wiAsqSPoWELDUdJXhjAhsVliSn7bWZjOhrgQ=
 github.com/go-admin-team/go-admin-core/v2 v2.5.0 h1:aD1SALklBxizGB9u8cOgm4OT8z656FM83F4fD6dMz9g=
 github.com/go-admin-team/go-admin-core/v2 v2.5.0/go.mod h1:LG/XvEfOplbuadKrPTPm0Nu5pN06aQUNZZC3ao4B4gs=
+github.com/go-admin-team/go-admin-core/v2 v2.6.0 h1:sRoZaxniTpbe287uR/uWpA14Jl1GTAcfGXLKoBLph2w=
+github.com/go-admin-team/go-admin-core/v2 v2.6.0/go.mod h1:LG/XvEfOplbuadKrPTPm0Nu5pN06aQUNZZC3ao4B4gs=
+github.com/go-admin-team/go-admin-core/v2 v2.7.0 h1:1qV0/5iFBvkE3BRtm4ip0v0QYG9Fgx4UtOTd8zkQT9c=
+github.com/go-admin-team/go-admin-core/v2 v2.7.0/go.mod h1:LG/XvEfOplbuadKrPTPm0Nu5pN06aQUNZZC3ao4B4gs=
 github.com/go-kit/kit v0.8.0/go.mod h1:xBxKIO96dXMWWy0MnWVtmwkA9/13aqxPnvrjFYMA2as=
 github.com/go-kit/kit v0.9.0/go.mod h1:xBxKIO96dXMWWy0MnWVtmwkA9/13aqxPnvrjFYMA2as=
 github.com/go-kit/kit v0.10.0/go.mod h1:xUsJbQ/Fp4kEt7AFgCuvyX4a71u8h9jB8tj/ORgOZ7o=