package cli

import (
	"context"
	"crypto/rand"
	"crypto/tls"
	"encoding/hex"
	"errors"
	"fmt"
	"log/slog"
	"maps"
	"net"
	"net/http"
	"os"
	"os/signal"
	"slices"
	"strings"
	"sync"
	"syscall"
	"time"

	"gitea.mynodes.xyz/ilker/llmgw/internal/clientip"
	"gitea.mynodes.xyz/ilker/llmgw/internal/control/inventory"
	"gitea.mynodes.xyz/ilker/llmgw/internal/control/modelserve"
	"gitea.mynodes.xyz/ilker/llmgw/internal/gguf"
	"gitea.mynodes.xyz/ilker/llmgw/internal/pgha"
	"gitea.mynodes.xyz/ilker/llmgw/internal/sourcecheck"
	"github.com/prometheus/client_golang/prometheus"

	"github.com/spf13/cobra"
	"golang.org/x/net/http2"
	"golang.org/x/net/http2/h2c"
	"golang.org/x/net/netutil"
	"golang.org/x/sync/errgroup"

	llmgwv1 "gitea.mynodes.xyz/ilker/llmgw/api/gen/llmgw/v1"
	"gitea.mynodes.xyz/ilker/llmgw/internal/audit"
	"gitea.mynodes.xyz/ilker/llmgw/internal/bench/eval"
	"gitea.mynodes.xyz/ilker/llmgw/internal/bundle"
	"gitea.mynodes.xyz/ilker/llmgw/internal/config"
	"gitea.mynodes.xyz/ilker/llmgw/internal/control"
	"gitea.mynodes.xyz/ilker/llmgw/internal/control/adminapi"
	"gitea.mynodes.xyz/ilker/llmgw/internal/control/agentsvc"
	"gitea.mynodes.xyz/ilker/llmgw/internal/control/catalog"
	"gitea.mynodes.xyz/ilker/llmgw/internal/control/fabrictest"
	"gitea.mynodes.xyz/ilker/llmgw/internal/control/federation"
	"gitea.mynodes.xyz/ilker/llmgw/internal/control/fleetmetrics"
	"gitea.mynodes.xyz/ilker/llmgw/internal/control/modelsvc"
	"gitea.mynodes.xyz/ilker/llmgw/internal/control/reconciler"
	"gitea.mynodes.xyz/ilker/llmgw/internal/guard"
	"gitea.mynodes.xyz/ilker/llmgw/internal/hf"
	"gitea.mynodes.xyz/ilker/llmgw/internal/kvevents"
	"gitea.mynodes.xyz/ilker/llmgw/internal/mirror"
	"gitea.mynodes.xyz/ilker/llmgw/internal/notify"
	"gitea.mynodes.xyz/ilker/llmgw/internal/obs"
	"gitea.mynodes.xyz/ilker/llmgw/internal/promq"
	"gitea.mynodes.xyz/ilker/llmgw/internal/router"
	"gitea.mynodes.xyz/ilker/llmgw/internal/secrets"
	storepkg "gitea.mynodes.xyz/ilker/llmgw/internal/store"
	"gitea.mynodes.xyz/ilker/llmgw/internal/tlsconf"
	"gitea.mynodes.xyz/ilker/llmgw/internal/version"
	"gitea.mynodes.xyz/ilker/llmgw/internal/web"
)

func newServerCmd() *cobra.Command {
	var (
		configPath    string
		listen        string
		adminListen   string
		metricsListen string
		endpoints     []string
	)

	cmd := &cobra.Command{
		Use:   "server",
		Short: "Run the router (phase 1; the control plane arrives in phase 3)",
		Long: "Serves the OpenAI API in front of a set of engine replicas.\n\n" +
			"In this phase the replica list is static: give it in the config file, or\n" +
			"pass --endpoint id=URL=model[,model] once per replica for a quick test.",
		RunE: func(cmd *cobra.Command, _ []string) error {
			log := logger()

			cfg, err := config.Load(configPath)
			if err != nil {
				return err
			}
			if listen != "" {
				cfg.Listen = listen
			}
			if adminListen != "" {
				cfg.AdminListen = adminListen
			}
			if metricsListen != "" {
				cfg.MetricsListen = metricsListen
			}
			for _, spec := range endpoints {
				ep, err := parseEndpointFlag(spec)
				if err != nil {
					return err
				}
				cfg.Endpoints = append(cfg.Endpoints, ep)
			}
			if err := cfg.Validate(); err != nil {
				return err
			}
			controlMode := cfg.ControlPlaneMode()
			if !controlMode && len(cfg.Endpoints) == 0 {
				return errors.New("nothing to serve: declare models with profiles, " +
					"or give --endpoint for a router-only run")
			}
			// Offline, nothing is asked of the internet: releases come from
			// the mirror's copy, and no registry is asked about newer weights.
			if cfg.Offline.Enabled() {
				cfg.Releases = cfg.ReleasesBase()
				cfg.Mirror.UpdateCheck = 0
			}

			// Before the router, so its tracer is the configured one. The
			// host tells two control planes' traces apart.
			host, _ := os.Hostname()
			shutdownTracing, err := obs.SetupTracing(cmd.Context(), obs.TracingOptions{
				Endpoint: cfg.Tracing.Endpoint, SampleRatio: cfg.Tracing.SampleRatio,
				Service: "llmgw-server", Version: version.Version, Instance: host,
			})
			if err != nil {
				return err
			}
			defer func() {
				// Bounded: a collector that is gone must not hold up a shutdown.
				sctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
				defer cancel()
				_ = shutdownTracing(sctx)
			}()

			rt, err := router.New(cfg, log)
			if err != nil {
				return err
			}
			// The store's own secrets sealed under a key kept outside it,
			// before anything reads them.
			if st := rt.Store(); st != nil && cfg.Secrets.MasterKey != "" {
				sealer, err := secrets.NewSealer(cmd.Context(), cfg.Secrets.MasterKey)
				if err != nil {
					return err
				}
				st.SetSealer(sealer)
			}
			// The longest any key may live, where every way of making one
			// reads it: the store.
			if st := rt.Store(); st != nil {
				if err := st.SetMaxKeyTTL(cmd.Context(), storepkg.ScopeGateway, "", cfg.Auth.MaxKeyTTL); err != nil {
					log.Warn("could not record auth.max_key_ttl", "err", err)
				}
			}
			// Guardrails, when server.yaml defines any. Attachments to
			// parts of the tenancy tree live in the store; without one, only
			// guard.default applies.
			var gate *guard.Gate
			if len(cfg.Guard.Policies) > 0 {
				var att guard.Attachments
				if st := rt.Store(); st != nil {
					att = st
				}
				if gate, err = guard.NewGate(cfg.Guard, att, log); err != nil {
					return fmt.Errorf("guard: %w", err)
				}
				gate.SetJudge(rt.Judge, rt.GuardFormat)
				rt.SetGuard(gate)
				rt.SetGuardLibrary(cfg.Guard)
				if err := rt.RegisterMetrics(gate); err != nil {
					return fmt.Errorf("guard metrics: %w", err)
				}
				log.Info("guardrails on", "policies", len(cfg.Guard.Policies), "default", cfg.Guard.Default)
			}

			ctx, stop := signal.NotifyContext(cmd.Context(), syscall.SIGINT, syscall.SIGTERM)
			defer stop()

			rt.Start(ctx)
			defer rt.Stop()

			// Which cluster this is, to federated peers: the id in the hop
			// header that stops a request going round two clusters that
			// federate each other, and the name in the tags a peer may be
			// sent. The id is the store's, so every control plane of one
			// cluster says the same.
			clusterID, err := federationID(ctx, rt.Store())
			if err != nil {
				return err
			}
			clusterName := cfg.Federation.Name
			if clusterName == "" {
				clusterName = host
			}
			rt.SetFederationIdentity(clusterID, clusterName)

			// Webhooks for what needs a person, when server.yaml names any.
			notifier, err := notify.New(ctx, cfg.Notify, log)
			if err != nil {
				return err
			}
			rt.SetNotifier(notifier)
			if gate != nil && notifier != nil {
				gate.SetJudgeAlarm(func(model string, err error) {
					notifier.Notify(notify.JudgeUnavailable, model,
						fmt.Sprintf("the guardrails judge %s gives no verdict (%v): policies with fail: closed "+
							"refuse what it should read, fail: open serve it unjudged", model, err),
						map[string]any{"model": model, "error": err.Error()})
				})
			}
			if notifier != nil {
				log.Info("notifications on", "webhooks", len(cfg.Notify.Webhooks))
			}

			// The control plane replaces the static endpoint list: agents
			// register, the reconciler places replicas, and Ready ones are
			// published to the router (plan §6.4).
			var agentHandler http.Handler
			var agentPath string
			var admin *control.Controller
			updateMetrics := &modelserve.UpdateCollector{}
			// The store's cluster, when it is one; read by the metrics and
			// by `ctl status`.
			var storeHA *pgha.Collector
			if controlMode {
				cat, err := cfg.Catalog()
				if err != nil {
					return err
				}
				live := catalog.NewLive(cat)
				// The router derives its own view, and it is empty until told.
				// Told here rather than at construction so there is exactly one
				// path by which it learns the catalog, used at startup and at
				// every change after.
				rt.SetCatalog(cat)

				ctl, svc, err := startControlPlane(ctx, cfg, live, rt, cfg.AgentTLS.Enabled(), log)
				if err != nil {
					return err
				}
				defer ctl.Close()
				ctl.Reconciler().SetNotifier(notifier)
				admin = ctl
				agentPath, agentHandler = svc.Handler()

				// Before AdminHandler, which is where the registry is built —
				// registering after it silently publishes nothing, which is
				// exactly what the first version of this did.
				//
				// On the same /metrics as the traffic figures: correlating
				// "the cluster was full" with "requests were queueing" should
				// not need two systems.
				// When the newest backup was written, so its absence alerts
				// rather than being found on the day it is needed.
				backupDir := cfg.BackupDir
				if backupDir == "" {
					backupDir = DefaultBackupDir
				}
				if err := rt.RegisterMetrics(prometheus.NewGaugeFunc(prometheus.GaugeOpts{
					Name: "llmgw_backup_last_success_timestamp_seconds",
					Help: "When the newest backup archive was written, as a Unix time; 0 when there is none.",
				}, func() float64 {
					if t := NewestBackup(backupDir); !t.IsZero() {
						return float64(t.Unix())
					}
					return 0
				})); err != nil {
					return err
				}
				if err := rt.RegisterMetrics(tlsconf.ExpiryCollector(map[string]string{
					"agent-server": cfg.AgentTLS.Cert, "agent-ca": cfg.AgentTLS.CA,
				})); err != nil {
					return err
				}
				if err := rt.RegisterMetrics(fleetmetrics.New(ctl)); err != nil {
					return err
				}
				// Whether models' repositories have new weights; given its
				// source once the mirror is set up, below.
				if err := rt.RegisterMetrics(updateMetrics); err != nil {
					return err
				}
				// A store of several hosts is a replicated cluster: what state
				// it is in is published beside everything else, read from its
				// Patroni. One host is a single PostgreSQL, which has none.
				if addr, err := storepkg.ResolveAddress(cfg.Auth.Store); err == nil && storepkg.IsPostgres(addr) {
					if hosts, _, _, err := storepkg.SplitPostgres(strings.TrimSpace(addr)); err == nil && len(hosts) > 1 {
						if cfg.AgentTLS.CA != "" {
							// Members' APIs, once on TLS, verified against it.
							if err := pgha.SetRESTCA(cfg.AgentTLS.CA); err != nil {
								log.Warn("the store's API cannot be verified", "err", err)
							}
						}
						storeHA = pgha.NewCollector(hosts)
						if err := rt.RegisterMetrics(storeHA); err != nil {
							return err
						}
					}
				}
				// So a request for a model that is configured but not running
				// starts it, rather than being told nothing serves it.
				rt.SetActivator(ctl.Reconciler())

				log.Info("control plane running",
					"models", len(cfg.Models), "agent_listen", cfg.AgentListen,
					"agent_mtls", cfg.AgentTLS.Enabled())
			}

			adminMux, err := rt.AdminHandler()
			if err != nil {
				return err
			}
			// Kept past this block: serving a model from the registry needs
			// both the admin API and the mirror, and the mirror is set up
			// below.
			var (
				adminAPI   *adminapi.API
				catalogSvc *modelsvc.Service
				evals      *eval.Service
			)
			if admin != nil {
				api := adminapi.New(admin)
				if storeHA != nil {
					api.SetStoreStatus(func() ([]adminapi.StoreMember, error) {
						c, err := storeHA.Read()
						var out []adminapi.StoreMember
						for _, m := range c.Members {
							out = append(out, adminapi.StoreMember{Name: m.Name, Host: m.Host, Role: m.Role, State: m.State, Lag: m.Lag})
						}
						return out, err
					})
				}
				if st := rt.Store(); st != nil {
					api.SetControlPlanes(newControlPlaneProbe(st, cfg.Listen).Status)
				}
				adminAPI = api
				api.SetReleases(cfg.Releases)
				api.SetNotifier(notifier)
				api.SetReplayer(rt.Replay)
				// Quality suites, asked of a model the way a replay is.
				evals = newEvalService(cfg, rt, admin, log)
				api.SetEval(evals)
				if st := rt.Store(); st != nil && gate != nil {
					api.SetGuard(&adminapi.GuardAdmin{Library: cfg.Guard, Store: st, Invalidate: gate.Invalidate})
				}
				if st := rt.Store(); st != nil {
					// Removing a node revokes its certificate.
					api.SetRevoker(st.RevokeNode)
					api.SetAudit(st)
				}
				// Catalog changes need somewhere to persist, so this is wired
				// only when there is a store. Without one, the endpoints say so
				// rather than accepting a model that vanishes on restart.
				if st := rt.Store(); st != nil {
					svc := modelsvc.New(cfg.Models, st, admin, log)
					svc.SetCheck(servableCheck(cfg.WantedEngines()["llamacpp"]))
					// Other llmgw clusters' models, offered as PEER/MODEL:
					// read before the catalog is built, so a model that falls
					// back to one is whole at startup.
					fed := federation.New(federation.Config{Self: clusterID, Interval: cfg.Federation.Interval},
						st, svc, admin.Catalog, log)
					if err := fed.Load(ctx); err != nil {
						log.Warn("could not read the federation peers; continuing without them", "err", err)
					}
					svc.SetFederated(fed.Models)
					admin.SetPeers(fed)
					api.SetFederation(fed)
					if err := rt.RegisterMetrics(fed); err != nil {
						return err
					}
					go fed.Run(ctx)
					// Rebuild the catalog from configuration plus what was
					// stored, and put that in force — otherwise models added
					// before the last restart would be silently absent until
					// someone added them again.
					if built, err := svc.Load(ctx); err != nil {
						log.Error("could not load stored models; continuing with configuration only", "err", err)
					} else {
						admin.ReplaceCatalog(ctx, built)
					}
					api.SetModels(svc)
					catalogSvc = svc
					// And the mirror page, so a model that has just been
					// fetched can be served without writing YAML by hand.
					rt.SetCatalogAdmin(svc)
					// Benchmark results live here too. A measurement costs real
					// GPU time; without somewhere to put it the number lives in
					// whoever ran it's terminal and is gone by morning.
					api.SetCapacity(st)
					// And what each group of the tenancy tree spent, which is
					// the other half of a quota being reviewable.
					api.SetUsage(st)
					// Capacity quotas, set and read over this API.
					api.SetQuotas(st)
				}
				api.Register(adminMux)

				// The dashboard lives here and only here. This listener is
				// bound to loopback by default and also serves pprof, which can
				// dump the process heap — where API keys sit in plaintext.
				// Reaching the page from elsewhere is an SSH tunnel's job:
				//   ssh -L 9090:127.0.0.1:9090 <host>
				var usage web.UsageSource
				if st := rt.Store(); st != nil {
					usage = st
				}
				// The dashboard's operational charts. llmgw publishes these
				// metrics; what it cannot do is remember them, so the history
				// is read back from wherever they were scraped to.
				//
				// Built before either copy of the page is registered, and given
				// to both. The first version built it afterwards and handed it
				// only to the one on the client port, so the copy an operator
				// reaches by SSH tunnel — the one on this very listener — had
				// no charts at all and said nothing about why.
				var history web.History
				if h := promq.New(cfg.Monitoring.Prometheus, cfg.Monitoring.Timeout); h != nil {
					history = h
					// And the console's charts, over the admin API.
					if adminAPI != nil {
						adminAPI.SetHistory(h)
					}
					log.Info("dashboard history configured",
						"prometheus", cfg.Monitoring.Prometheus)
				}
				rt.SetHistory(history, cfg.Monitoring.Window)

				adminMux.HandleFunc("GET /", web.Dashboard(
					api, usage, version.Version, 5*time.Second, true,
					web.NavExtras{History: history, Window: cfg.Monitoring.Window}, log))
				adminMux.Handle("GET /_static/", web.Static())
				adminMux.HandleFunc("GET /_lang/{code}", web.SwitchLanguage)
				adminMux.HandleFunc("GET /nodes/{id}/logs", web.NodeLogsPage(api, true, version.Version))
				adminMux.HandleFunc("GET /nodes/{id}", web.NodePage(api, history, cfg.Monitoring.Window, true, version.Version, log))
				adminMux.HandleFunc("GET /nvlink", web.NVLinkPage(api, true, version.Version))
				adminMux.HandleFunc("GET /fabrics", web.FabricsPage(api, true, version.Version))
				adminMux.HandleFunc("GET /fabrics/{id}", web.FabricsPage(api, true, version.Version))
				adminMux.HandleFunc("GET /fabrics/{id}/topology.json", web.FabricTopologyJSON(api))

				// The same page on the client-facing port at /admin, behind a
				// session and an admin flag. Handing the router the data rather
				// than the port keeps pprof where it is.
				rt.SetClusterView(api, usage)
			}

			// The model mirror, when one is configured. Built here and not at
			// first use so a mistyped address stops the control plane starting,
			// rather than turning into a broken admin page hours later when
			// somebody finally opens it.
			if cfg.Mirror.API != "" {
				mc, err := mirror.NewClient(cfg.Mirror.API, cfg.Mirror.TLS, cfg.Mirror.Timeout)
				if err != nil {
					return fmt.Errorf("the model mirror: %w", err)
				}
				rt.SetMirror(mc)
				// Browsing the registry is what turns "paste a resolve URL"
				// into "find the model". It needs a way out to the internet
				// from this machine, which the mirror's own fetching does not.
				registry := hf.New(cfg.Mirror.Registry, cfg.Mirror.Timeout)
				rt.SetRegistry(registry)
				// One command from a registry repository to a replica that
				// answers: `llmgw ctl model serve`.
				if adminAPI != nil && catalogSvc != nil {
					serve := modelserve.New(modelserve.Deps{
						Registry: registry, Mirror: mc, Catalog: catalogSvc, Fleet: admin,
						LlamaVersion: cfg.WantedEngines()["llamacpp"], Log: log,
						// Updates: the catalog in force, the service that
						// changes it, and what replaces replicas.
						Current: func() []catalog.Model { return admin.Catalog().Models() },
						Models:  catalogSvc, Restart: admin.Reconciler().Restart,
						Pool: adminAPI.PoolNodes, Notify: notifier,
						// Canaries: how the router saw each side serve.
						CanaryStats: rt.CanaryStats, ResetCanary: rt.ResetCanary,
						// Updates and promotions held to a quality suite.
						Eval: evals,
					})
					go serve.WatchCanaries(ctx, 30*time.Second)
					adminAPI.SetServe(serve)
					updateMetrics.Set(serve)
					go serve.WatchUpdates(ctx, cfg.Mirror.UpdateCheck)
				}
				log.Info("model mirror configured",
					"api", cfg.Mirror.API, "mtls", cfg.Mirror.TLS.Enabled())
			}

			apiSrv := withLimits(&http.Server{
				Handler: rt.Handler(),
				// Reading the request: long enough for a body up to
				// max_body_bytes on a slow link, short enough that a client
				// trickling one cannot hold a connection and its buffer.
				ReadTimeout: 5 * time.Minute,
				// No write timeout: responses stream for minutes and the proxy
				// enforces its own per-phase deadlines.
			})
			if err := checkAdminListen(cfg); err != nil {
				return err
			}
			if bad := clientip.SetTrusted(cfg.Router.TrustedProxies); len(bad) > 0 {
				return fmt.Errorf("router.trusted_proxies: %q is not an address or a CIDR prefix", bad[0])
			}
			adminToken, err := ensureAdminToken(adminTokenPath(configPath, cfg))
			if err != nil {
				return err
			}
			adminHandler := auditAPI(rt, adminMux, log)
			// With a rival, a standby's admin API is the leader's: what the
			// console and ctl show lives in the leader's memory.
			var forward *leaderForward
			if admin != nil && admin.Leadership() != nil && rt.Store() != nil && cfg.AgentTLS.Enabled() {
				self, _ := os.Hostname()
				forward = &leaderForward{st: rt.Store(), lead: admin.Leadership(), self: self, tls: cfg.AgentTLS, log: log}
			}
			served := adminHandler
			if forward != nil {
				served = forward.standby(adminHandler)
			}
			adminSrv := withLimits(&http.Server{
				Handler:     router.AdminGuard(router.AdminAuth(adminToken, cfg.DebugPprof, served)),
				ReadTimeout: time.Minute,
			})

			apiLn, err := net.Listen("tcp", cfg.Listen)
			if err != nil {
				return fmt.Errorf("listening on %s: %w", cfg.Listen, err)
			}
			if cfg.MaxConnections > 0 {
				// Past this, new connections wait in the kernel's backlog
				// rather than each taking a goroutine and its buffers.
				apiLn = netutil.LimitListener(apiLn, cfg.MaxConnections)
			}
			adminLn, err := net.Listen("tcp", cfg.AdminListen)
			if err != nil {
				_ = apiLn.Close()
				return fmt.Errorf("listening on %s: %w", cfg.AdminListen, err)
			}

			// A metrics-only listener, when one is configured. Started after
			// AdminHandler, which is where the registry it serves is built.
			var metricsSrv *http.Server
			var metricsLn net.Listener
			if cfg.MetricsListen != "" {
				h := rt.MetricsHandler()
				if h == nil {
					_ = apiLn.Close()
					_ = adminLn.Close()
					return fmt.Errorf("metrics handler is not ready; this is a wiring fault, not a configuration one")
				}
				metricsLn, err = net.Listen("tcp", cfg.MetricsListen)
				if err != nil {
					_ = apiLn.Close()
					_ = adminLn.Close()
					return fmt.Errorf("listening on %s: %w", cfg.MetricsListen, err)
				}
				// Beside the metrics, the machines to scrape for their node
				// exporters, for Prometheus' http_sd_configs: on this port
				// because Prometheus already reaches it.
				if adminAPI != nil {
					mux := http.NewServeMux()
					mux.HandleFunc("GET /prometheus/targets", adminAPI.PromSD)
					mux.Handle("/", h)
					h = mux
				}
				metricsSrv = withLimits(&http.Server{Handler: h, ReadTimeout: 30 * time.Second})
				go func() {
					if err := metricsSrv.Serve(metricsLn); err != nil && !errors.Is(err, http.ErrServerClosed) {
						log.Error("metrics listener stopped", "err", err)
					}
				}()
			}

			log.Info("router listening",
				"api", apiLn.Addr().String(),
				"admin", adminLn.Addr().String(),
				"metrics", metricsAddr(metricsLn),
				"endpoints", len(cfg.Endpoints),
				"default_profile", cfg.Router.DefaultProfile,
			)

			var agentSrv *http.Server
			var agentLn net.Listener
			if agentHandler != nil {
				mux := http.NewServeMux()
				mux.Handle(agentPath, agentHandler)
				if forward != nil {
					mux.Handle(forwardPrefix+"/", router.AdminGuard(forward.leaderSide(adminHandler)))
				}
				// No read timeout: an agent's stream is one request that
				// lasts as long as the agent does.
				agentSrv = withLimits(&http.Server{Handler: mux})
				if !cfg.AgentTLS.Enabled() {
					// Without TLS, agents speak plaintext HTTP/2, which Go will
					// not do unless asked. Nothing then stops a node from
					// claiming to be another one, so this belongs on a trusted
					// network only.
					agentSrv.Handler = h2c.NewHandler(mux, &http2.Server{})
				}
				agentAddr, aerr := agentListenAddr(cfg)
				if aerr != nil {
					_ = apiLn.Close()
					_ = adminLn.Close()
					return aerr
				}
				if !cfg.AgentTLS.Enabled() {
					log.Warn("the agent port has no TLS (agent_insecure): agents are not authenticated and "+
						"traffic is plaintext; use only on one machine or a trusted network", "listen", agentAddr)
				}
				if agentLn, err = net.Listen("tcp", agentAddr); err != nil {
					_ = apiLn.Close()
					_ = adminLn.Close()
					return fmt.Errorf("listening on %s: %w", agentAddr, err)
				}
				if cfg.AgentTLS.Enabled() {
					tc, terr := cfg.AgentTLS.ServerConfig()
					if terr != nil {
						_ = apiLn.Close()
						_ = adminLn.Close()
						_ = agentLn.Close()
						return terr
					}
					// Serve on an already-wrapped listener rather than
					// ServeTLS: agentSrv.TLSConfig stays nil, which is exactly
					// the case where net/http sets up HTTP/2 for ALPN itself.
					agentLn = tls.NewListener(agentLn, tc)
				}
			}

			g, gctx := errgroup.WithContext(ctx)
			g.Go(func() error { return serve(apiSrv, apiLn) })
			g.Go(func() error { return serve(adminSrv, adminLn) })
			if agentSrv != nil {
				g.Go(func() error { return serve(agentSrv, agentLn) })
			}
			g.Go(func() error {
				<-gctx.Done()
				log.Info("shutting down")
				sctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
				defer cancel()
				if agentSrv != nil {
					// Closed, not shut down gracefully: an agent's stream
					// never goes idle, so Shutdown would wait out its whole
					// timeout while every node stayed on a control plane
					// that is leaving, and a standby could not be found
					// until it gave up. Closed, they move at once.
					_ = agentSrv.Close()
				}
				if metricsSrv != nil {
					_ = metricsSrv.Shutdown(sctx)
				}
				_ = adminSrv.Shutdown(sctx)
				return apiSrv.Shutdown(sctx)
			})
			return g.Wait()
		},
	}

	f := cmd.Flags()
	f.StringVar(&configPath, "config", "", "configuration file (optional)")
	f.StringVar(&listen, "listen", "", "override the client-facing address")
	f.StringVar(&adminListen, "admin-listen", "", "override the dashboard/pprof address")
	f.StringVar(&metricsListen, "metrics-listen", "", "serve metrics only on this address, safe to expose")
	f.StringArrayVar(&endpoints, "endpoint", nil, "static endpoint: id=URL=model[,model] (repeatable)")

	cmd.AddCommand(newServerInitCmd(), newServerAddCmd())
	cmd.AddCommand(newBackupCmd(), newRestoreCmd())
	return cmd
}

// startControlPlane builds the control plane and its agent service.
//
// The service is both the AgentService handler and the reconciler's Dispatcher:
// it holds the open streams, so it is the only thing that can deliver a command
// to a node.
func startControlPlane(ctx context.Context, cfg config.Server, cat *catalog.Live,
	sink control.EndpointSink, requireIdentity bool, log *slog.Logger,
) (*control.Controller, *agentsvc.Service, error) {
	ccfg := control.DefaultConfig()
	ccfg.Reconciler = reconciler.Config{
		Interval:         cfg.Control.Interval,
		BaseBackoff:      cfg.Control.BaseBackoff,
		MaxBackoff:       cfg.Control.MaxBackoff,
		UnreachableAfter: cfg.Control.UnreachableAfter,
		LostAfter:        cfg.Control.LostAfter,
		MgmtInterface:    cfg.Control.MgmtInterface,
	}
	if cfg.Control.Timezone != "" {
		loc, err := time.LoadLocation(cfg.Control.Timezone)
		if err != nil {
			return nil, nil, fmt.Errorf("control.timezone: %w", err)
		}
		ccfg.Reconciler.Location = loc
	}
	if len(cfg.Control.PortRange) == 2 {
		ccfg.Reconciler.PortRange = [2]int{cfg.Control.PortRange[0], cfg.Control.PortRange[1]}
	}

	// Each replica's engine demands a key only the router has. The secret it
	// is derived from lives in the store, so every control plane against it —
	// and this one after a restart — computes the same key for a replica.
	if h, ok := sink.(interface{ Store() *storepkg.DB }); ok && h.Store() != nil {
		secret, err := h.Store().ClusterSecret(ctx, "engine-key")
		if err != nil {
			return nil, nil, fmt.Errorf("reading the engine key secret: %w", err)
		}
		ccfg.Reconciler.EngineKey = control.EngineKeys([]byte(secret))
	} else {
		log.Warn("engines start without a key: there is no store to keep the secret " +
			"they are derived from, so anyone who can reach a node's ports reaches its models")
	}

	// The service needs the inventory and the status sink, both of which the
	// controller owns; the controller needs the service as its dispatcher. The
	// cycle is broken by giving the controller a dispatcher that forwards to
	// the service once it exists.
	fwd := &dispatcherRef{}
	ctl := control.New(ccfg, cat, fwd, sink, log)
	// The router is both where endpoints go and where activity comes back
	// from. It only reports usage when it is the thing serving requests, which
	// is why this is a type assertion rather than a parameter.
	if u, ok := sink.(control.UsageSource); ok {
		ctl.SetUsageSource(u)
	}
	// Power caps and estimates; Validate has already refused a bad one.
	if pol, err := cfg.Facility.Policy(); err == nil {
		ctl.Facility().SetPolicy(pol)
	}
	// With a shared store, two control planes can see each other, so this one
	// competes rather than assuming it is alone. Without one there is nobody
	// to compete with and it places unconditionally, exactly as before.
	// The same shape as the usage assertion above, and for the same reason:
	// the sink is the router when the router is serving, and nothing when a
	// control plane runs on its own.
	var lead *control.Leadership
	if h, ok := sink.(interface{ Store() *storepkg.DB }); ok {
		if st := h.Store(); st != nil {
			lead = control.NewLeadership(st, cfg.Control.LeaseTTL, log)
			if r, ok := sink.(interface{ SetLeader(func() bool) }); ok {
				r.SetLeader(lead.IsLeader)
			}
			ctl.SetLeadership(lead)
			// The tenancy tree's capacity quotas: what each part is
			// guaranteed and may hold.
			ctl.Reconciler().SetQuotas(control.QuotaPolicy(st, log))
			// And what each model was measured to serve, for scaling and for
			// choosing between profiles.
			ctl.Reconciler().SetMeasurements(control.Measurements(st, log))
			// The engine builds chosen per node on the cluster page.
			ctl.SetEngineChoices(control.NewEngineChoices(ctx, st, log))
			// Domains held for maintenance and labels set on nodes.
			ctl.Facility().UseStore(ctx, st)
			log.Info("competing to be the control plane that places work",
				"owner", lead.Owner())
		}
	}
	svc := agentsvc.New(ctl.Inventory(), ctl.Reconciler(), requireIdentity, log)
	svc.SetEnginesFor(func(n inventory.Node) map[string]string {
		return cfg.EnginesFor(n.ID, slices.Sorted(maps.Keys(n.GPUsByClass())), n.Labels)
	})
	// The KV cache events agents forward go to the router's precise index,
	// when there is a router in this process to have one.
	if k, ok := sink.(interface{ ApplyKV(kvevents.Update) }); ok {
		svc.SetKVSink(k.ApplyKV)
	}
	log.Info("nodes install these engines", "engines", cfg.WantedEngines(), "overrides", len(cfg.EngineOverrides))
	if cfg.Offline.Enabled() {
		// From the mirror, each file held to the checksum the bundle gave,
		// which this control plane keeps and the mirror has no say in.
		trusted := bundle.NewTrusted(cfg.Offline.ChecksumsFile())
		svc.SetOffline(func(engine, version string) (string, map[string]string) {
			return bundle.EngineBase(cfg.Offline.Mirror, engine, version), trusted.For(engine, version)
		})
		if v, ok := cfg.WantedEngines()["llamacpp"]; ok && len(trusted.For("llamacpp", v)) == 0 {
			log.Warn("offline, and no bundle imported here carries llama.cpp at this version: nodes cannot "+
				"install it until one does (llmgw bundle create --engine llamacpp="+v+", then bundle import or trust)",
				"version", v, "checksums", cfg.Offline.ChecksumsFile(), "err", trusted.Err())
		}
		log.Info("offline: nodes install engines from the mirror", "mirror", cfg.Offline.Mirror,
			"releases", cfg.Releases)
	}
	if ch := ctl.EngineChoices(); ch != nil {
		svc.SetEngineChoices(ch.For)
		ch.OnChange(svc.Reprovision)
	}
	// A standby must refuse agents as well as refusing to place. Commands
	// travel down the stream the agent opened, so a standby holding one is a
	// node the leader cannot command — and the leader's placements would fail
	// with "no agent connected" for a node that is plainly connected.
	if lead != nil {
		svc.SetLeadership(lead)
	}
	// The control planes agents may dial, from the store every control plane
	// shares: told to each agent as it registers, and to all of them when
	// `server add` records another.
	if h, ok := sink.(interface{ Store() *storepkg.DB }); ok && h.Store() != nil {
		st := h.Store()
		svc.SetControlPlanes(func(ctx context.Context) ([]string, error) {
			cps, err := st.ControlPlanes(ctx)
			urls := make([]string, len(cps))
			for i, c := range cps {
				urls[i] = c.URL
			}
			return urls, err
		})
		go svc.WatchControlPlanes(ctx, 30*time.Second)
		// A removed node's certificate is refused.
		svc.SetRevocations(st.NodeRevokedAt)
	}
	fwd.set(svc)
	// Every InfiniBand fabric mapped every five minutes, through one node.
	go svc.ScanFabrics(ctx, 5*time.Minute)

	ctl.Start(ctx)
	return ctl, svc, nil
}

// auditAPI records every change made through the admin API, when there is a
// store to record it in.
func auditAPI(rt *router.Router, next http.Handler, log *slog.Logger) http.Handler {
	st := rt.Store()
	if st == nil {
		return next
	}
	return audit.API(st, adminapi.Actor, log, next)
}

// metricsAddr renders the metrics listener for the startup line, saying so
// plainly when there is none rather than printing an empty string.
func metricsAddr(ln net.Listener) string {
	if ln == nil {
		return "off (admin listener only)"
	}
	return ln.Addr().String()
}

// dispatcherRef lets the controller be built before the service exists.
type dispatcherRef struct {
	mu sync.RWMutex
	d  reconciler.Dispatcher
}

func (r *dispatcherRef) set(d reconciler.Dispatcher) {
	r.mu.Lock()
	r.d = d
	r.mu.Unlock()
}

func (r *dispatcherRef) get() (reconciler.Dispatcher, error) {
	r.mu.RLock()
	defer r.mu.RUnlock()
	if r.d == nil {
		return nil, errors.New("the agent service is not ready yet")
	}
	return r.d, nil
}

// RemoveWeights forwards a capability beyond scheduling.
//
// It has to be spelled out here rather than inherited, because this type
// satisfies reconciler.Dispatcher by forwarding each method by hand — so
// anything the controller reaches by type assertion stops at this boundary
// unless it is forwarded too. A capability added to the service and not to
// this type is simply absent at runtime, with nothing at compile time to say
// so.
func (r *dispatcherRef) RemoveWeights(ctx context.Context, nodeID string, paths []string) error {
	d, err := r.get()
	if err != nil {
		return err
	}
	rm, ok := d.(control.WeightRemover)
	if !ok {
		return errors.New("the agent service cannot remove weights")
	}
	return rm.RemoveWeights(ctx, nodeID, paths)
}

// Logs forwards a node log request, the same way RemoveWeights does: the
// controller finds it by type assertion, so a forwarder without it would make
// the capability silently absent at runtime.
func (r *dispatcherRef) Logs(ctx context.Context, nodeID, source string, lines int) ([]string, []string, error) {
	d, err := r.get()
	if err != nil {
		return nil, nil, err
	}
	lr, ok := d.(control.LogReader)
	if !ok {
		return nil, nil, errors.New("this dispatcher cannot read node logs")
	}
	return lr.Logs(ctx, nodeID, source, lines)
}

func (r *dispatcherRef) FabricTest(ctx context.Context, nodeID string, req *llmgwv1.FabricTestRequest) (*llmgwv1.FabricTestReply, error) {
	d, err := r.get()
	if err != nil {
		return nil, err
	}
	ft, ok := d.(fabrictest.Asker)
	if !ok {
		return nil, errors.New("this dispatcher cannot run a bandwidth test")
	}
	return ft.FabricTest(ctx, nodeID, req)
}

// Upgrade forwards a node upgrade, as Logs does.
func (r *dispatcherRef) Upgrade(ctx context.Context, nodeID, version, releases string) error {
	d, err := r.get()
	if err != nil {
		return err
	}
	u, ok := d.(control.Upgrader)
	if !ok {
		return errors.New("this dispatcher cannot upgrade nodes")
	}
	return u.Upgrade(ctx, nodeID, version, releases)
}

func (r *dispatcherRef) StartShard(ctx context.Context, cmd reconciler.ShardCommand) error {
	d, err := r.get()
	if err != nil {
		return err
	}
	return d.StartShard(ctx, cmd)
}

func (r *dispatcherRef) StopShard(ctx context.Context, nodeID, shardID string) error {
	d, err := r.get()
	if err != nil {
		return err
	}
	return d.StopShard(ctx, nodeID, shardID)
}

func (r *dispatcherRef) DrainShard(ctx context.Context, nodeID, shardID string) error {
	d, err := r.get()
	if err != nil {
		return err
	}
	return d.DrainShard(ctx, nodeID, shardID)
}

func serve(s *http.Server, ln net.Listener) error {
	if err := s.Serve(ln); err != nil && !errors.Is(err, http.ErrServerClosed) {
		return err
	}
	return nil
}

// parseEndpointFlag decodes id=URL=model[,model].
//
// The URL contains '=' only in pathological cases, so splitting into exactly
// three parts keeps the flag readable while staying unambiguous.
func parseEndpointFlag(spec string) (config.Endpoint, error) {
	parts := strings.SplitN(spec, "=", 3)
	if len(parts) != 3 {
		return config.Endpoint{}, fmt.Errorf("--endpoint %q: expected id=URL=model[,model]", spec)
	}
	models := strings.Split(parts[2], ",")
	for i := range models {
		models[i] = strings.TrimSpace(models[i])
	}
	return config.Endpoint{
		ID:     strings.TrimSpace(parts[0]),
		URL:    strings.TrimRight(strings.TrimSpace(parts[1]), "/"),
		Models: models,
		Pool:   "static",
		Weight: 1,
	}, nil
}

// servableCheck refuses a model whose llama.cpp weights the nodes' build
// cannot load, read from the weights' own header before anything is placed.
func servableCheck(llamaVersion string) func(context.Context, catalog.Model) error {
	return func(ctx context.Context, m catalog.Model) error {
		if llamaVersion == "" {
			return nil
		}
		usesLlama := false
		for _, p := range m.Profiles {
			if p.Engine == "llamacpp" {
				usesLlama = true
			}
		}
		url, ok := sourcecheck.HTTPSource(m.Source)
		if !usesLlama || !ok || !gguf.IsGGUF(url) {
			return nil
		}
		if v := gguf.Check(ctx, url, llamaVersion); !v.Servable {
			return fmt.Errorf("%s cannot be served: %s", m.Name, v.Reason)
		}
		return nil
	}
}

// errAgentInsecure is a control plane asked to open its agent port with no
// TLS and not told that is meant.
var errAgentInsecure = errors.New("agent_tls is not set, so the agent port would be plaintext and anyone " +
	"reaching it could register as a node: set agent_tls (llmgw server init writes it), or " +
	"agent_insecure: true for one machine or a test")

// agentListenAddr is where the agent port listens: agent_listen with TLS;
// without it, refused unless agent_insecure says so, and then on loopback
// when agent_listen names only a port.
func agentListenAddr(cfg config.Server) (string, error) {
	if cfg.AgentTLS.Enabled() {
		return cfg.AgentListen, nil
	}
	// agent_tls.insecure says the same, in the older place.
	if !cfg.AgentInsecure && !cfg.AgentTLS.Insecure {
		return "", errAgentInsecure
	}
	if host, port, err := net.SplitHostPort(cfg.AgentListen); err == nil && host == "" {
		return net.JoinHostPort("127.0.0.1", port), nil
	}
	return cfg.AgentListen, nil
}

// withLimits gives a server what every listener here wants: a bound on how
// long the headers may take, how long a connection may sit idle between
// requests, and how large the headers may be.
func withLimits(s *http.Server) *http.Server {
	s.ReadHeaderTimeout = 10 * time.Second
	s.IdleTimeout = 2 * time.Minute
	s.MaxHeaderBytes = 1 << 20
	return s
}

// federationID is this cluster's id among federated peers: kept in the store,
// so every control plane of the cluster and every restart says the same. A
// gateway with no store makes one for its lifetime, which is all a router
// with no peers of its own needs to refuse a request that comes back to it.
func federationID(ctx context.Context, st *storepkg.DB) (string, error) {
	if st == nil {
		b := make([]byte, 8)
		if _, err := rand.Read(b); err != nil {
			return "", err
		}
		return hex.EncodeToString(b), nil
	}
	secret, err := st.ClusterSecret(ctx, "cluster-id")
	if err != nil {
		return "", fmt.Errorf("this cluster's id for federation: %w", err)
	}
	// A prefix is plenty to tell clusters apart, and the id is sent to
	// peers: nothing is derived from it, but there is no reason to send
	// more of a secret than needed.
	return secret[:16], nil
}