// Command specsrht is the spec.sr.ht daemon, and — invoked under another name
// — the receive hooks of every space it serves.
//
// It runs three things in one process:
//
//   - the hook RPC socket at <repos>/.specsrht/hook.sock, which the pre-receive,
//     update and post-receive hooks of every space call. Validation lives here
//     and not in the hooks because bleve is single-writer and this process holds
//     the index, and because the push path and the API must not be able to
//     disagree about what is valid.
//   - an HTTP listener on -b (default localhost:5091), serving /healthz, the
//     read UI, /mcp for agents and /query for GraphQL — one listener, so one
//     reverse-proxy route covers the service.
//   - the reconciler, at startup and then periodically, repairing the
//     divergence a killed daemon leaves between git refs, Postgres and the
//     index.
//
// # Running as a hook
//
// Every hook a space's repository runs is a symlink to this binary. When
// argv[0] names a hook the process handles that hook and exits, touching
// neither the configuration nor the database; see the hooks package. The
// explicit form `specsrht hook <name>` does the same thing by hand.
//
// # Flags
//
// Parsed by core-go's server.New:
//
//	-b addr   bind address (repeatable); default localhost:5091
//	-d        debug (verbose request logging in core-go)
//	-m addr   Prometheus metrics bind (default random port)
//	-p addr   pprof bind (default random localhost port)
//
// Configuration comes from the shared SourceHut config.ini through core-go's
// fixed search path. Every required key is checked before anything is opened,
// and all missing ones are reported together, so a misconfigured instance
// fails once at startup rather than once per restart.
//
// # Shutdown
//
// SIGINT and SIGTERM both start a warm shutdown. core-go's server.Run only
// handles SIGINT — which is why compare.sr.ht's systemd unit sets
// KillSignal=SIGINT — so this daemon installs a handler that turns a SIGTERM
// into that same SIGINT. A unit for this service therefore needs no
// KillSignal= line: the systemd default works.
package main

import (
	"context"
	"database/sql"
	"errors"
	"fmt"
	"log/slog"
	"net/http"
	"os"
	"os/signal"
	"path/filepath"
	"strings"
	"syscall"
	"time"

	"github.com/go-chi/chi/v5"
	chimw "github.com/go-chi/chi/v5/middleware"
	_ "github.com/lib/pq" // registers the "postgres" database/sql driver
	"github.com/vaughan0/go-ini"
	"go.bigb.es/auxilia/scribe"

	"sourcecraft.dev/bigbes/sr-ht-core/config"
	coreserver "sourcecraft.dev/bigbes/sr-ht-core/server"

	"sourcecraft.dev/bigbes/sr-ht-spec/core"
	"sourcecraft.dev/bigbes/sr-ht-spec/graph"
	"sourcecraft.dev/bigbes/sr-ht-spec/hooks"
	"sourcecraft.dev/bigbes/sr-ht-spec/mcpsrv"
	"sourcecraft.dev/bigbes/sr-ht-spec/search"
	"sourcecraft.dev/bigbes/sr-ht-spec/service"
	"sourcecraft.dev/bigbes/sr-ht-spec/web"
)

const (
	// serviceName is the SourceHut service identifier and our config section.
	// The ".sr.ht" suffix is what puts us in the nav network list.
	serviceName = "spec.sr.ht"

	// defaultBind is the address core-go binds when no -b is given.
	defaultBind = "localhost:5091"

	// version is reported in the MCP handshake so a client listing several
	// SourceHut MCP endpoints can tell which build it is talking to.
	version = "dev"

	// pingTimeout bounds the startup connectivity check against Postgres.
	pingTimeout = 10 * time.Second

	// shutdownGrace is how long the hook socket is given to finish the calls
	// already in flight once the HTTP listener has drained. A push being
	// validated at that moment finishes rather than being failed closed.
	shutdownGrace = 30 * time.Second
)

func main() {
	// Hook mode first: a hook must not read a config file, open Postgres, or
	// bind anything. It talks to the daemon over a socket and exits.
	if _, _, isHook := hooks.ModeFromArgs(os.Args); isHook {
		os.Exit(hooks.Run(hooks.Runtime{Args: os.Args}))
	}

	log := newLogger()
	slog.SetDefault(log)

	if err := run(log); err != nil {
		// Plain text, not a log record. A startup failure is read by a human
		// on a terminal, and the configuration report is deliberately several
		// lines long — a structured handler would escape it into one.
		fmt.Fprintf(os.Stderr, "spec.sr.ht did not start: %v\n", err)
		os.Exit(1)
	}
}

// newLogger builds the process logger. LOG_LEVEL raises or lowers verbosity;
// everything goes to stderr, because a hook's stdout is forwarded to the
// pushing client and this binary is both programs.
func newLogger() *slog.Logger {
	level := new(slog.LevelVar)
	level.Set(parseLevel(os.Getenv("LOG_LEVEL")))
	return slog.New(scribe.NewTintHandler(
		scribe.WithWriter(os.Stderr),
		scribe.WithLevel(level),
	))
}

func parseLevel(s string) slog.Level {
	switch strings.ToLower(strings.TrimSpace(s)) {
	case "debug":
		return slog.LevelDebug
	case "warn", "warning":
		return slog.LevelWarn
	case "error":
		return slog.LevelError
	default:
		return slog.LevelInfo
	}
}

func run(log *slog.Logger) error {
	// LoadConfig never fails on a missing file — it returns a nil ini.File —
	// so validateConfig is what turns an unconfigured instance into one clear
	// message instead of a panic deep inside the first request.
	conf := config.LoadConfig()
	cfg, err := validateConfig(conf)
	if err != nil {
		return err
	}

	pool, err := openDatabase(cfg.ConnectionString)
	if err != nil {
		return err
	}
	defer pool.Close()

	svc, err := service.New(cfg, pool)
	if err != nil {
		return err
	}

	// Refresh every space's hooks before anything can be pushed to it. This is
	// fatal on failure by design: a space whose hooks are missing accepts
	// pushes that are never validated, which is the one outcome the whole
	// receive path exists to prevent. A daemon that will not start is loud; a
	// space quietly accepting malformed documents is not.
	binary, err := os.Executable()
	if err != nil {
		return fmt.Errorf("locate this binary, which every hook symlinks to: %w", err)
	}
	if err := refreshHooks(context.Background(), log, svc, binary); err != nil {
		return err
	}

	hookSrv, err := hooks.NewServer(hooks.Options{
		Backend: svc,
		Socket:  hooks.SocketPath(cfg.Repos),
		Log:     log,
		OnPush:  pushNotifier(log),
	})
	if err != nil {
		return err
	}
	if err := hookSrv.Listen(); err != nil {
		return err
	}

	surf, err := newSurfaces(conf, cfg, svc, version)
	if err != nil {
		return err
	}
	defer surf.Close()

	// server.New parses -b/-d/-m/-p and runs crypto.InitCrypto(conf), whose
	// two required keys validateConfig already checked, so it cannot fatal
	// here for a reason we have not already reported.
	srv := coreserver.New(serviceName, defaultBind, conf, os.Args)
	mountRoutes(srv.AnonRouter(), conf, surf)

	ctx, stop := context.WithCancel(context.Background())
	defer stop()

	served := make(chan error, 1)
	go func() { served <- hookSrv.Serve(ctx) }()
	go svc.RunReconciler(ctx, service.DefaultReconcileInterval, reconcileReporter(log))

	bridgeSIGTERM(log)

	log.Info("spec.sr.ht starting",
		"bind", defaultBind,
		"repos", cfg.Repos,
		"cache", cfg.Cache,
		"origin", cfg.Origin,
		"hook_socket", hookSrv.Socket(),
		"reconcile_interval", service.DefaultReconcileInterval.String(),
	)

	// Blocks until SIGINT — which bridgeSIGTERM makes SIGTERM equivalent to —
	// and then drains the HTTP listeners.
	srv.Run()

	log.Info("draining the hook socket", "grace", shutdownGrace.String())
	stop()
	select {
	case err := <-served:
		if err != nil {
			log.Error("hook socket stopped with an error", scribe.Err(err))
		}
	case <-time.After(shutdownGrace):
		log.Warn("hook socket did not drain in time; closing it")
	}
	if err := hookSrv.Close(); err != nil {
		log.Error("could not close the hook socket", scribe.Err(err))
	}
	log.Info("spec.sr.ht stopped")
	return nil
}

// validateConfig checks every key this daemon needs before anything is opened,
// and reports all of the missing ones at once so an operator fixes the config
// in one pass instead of discovering each gap on a separate restart.
//
// service.LoadConfig owns our own section and collects its own gaps the same
// way; the two lists are merged into one message. The keys checked here are
// the ones core-go itself fatals on, which belong to server.New's contract
// rather than to service/ — duplicating them there would give the instance two
// lists to keep in sync.
func validateConfig(conf ini.File) (service.Config, error) {
	var missing []string
	require := func(section, key, why string) {
		if v, ok := conf.Get(section, key); !ok || strings.TrimSpace(v) == "" {
			missing = append(missing, fmt.Sprintf("[%s] %s — %s", section, key, why))
		}
	}

	// Both are read by crypto.InitCrypto, which server.New calls and which
	// fatals with a terse message when either is absent. The webhook key is
	// required even though v1 emits no webhooks.
	require("sr.ht", "network-key", "fernet key for the unified-login cookie")
	require("webhooks", "private-key", "webhook signing key; crypto.InitCrypto requires it")

	cfg, cfgErr := service.LoadConfig(conf)

	if len(missing) == 0 && cfgErr == nil {
		return cfg, nil
	}

	var b strings.Builder
	b.WriteString("incomplete configuration.")
	if len(missing) > 0 {
		fmt.Fprintf(&b, "\n\nMissing keys the SourceHut runtime requires:\n\t%s",
			strings.Join(missing, "\n\t"))
	}
	if cfgErr != nil {
		fmt.Fprintf(&b, "\n\n%s", cfgErr)
	}
	return service.Config{}, errors.New(b.String())
}

// openDatabase opens the pool the whole daemon shares — request handlers, the
// reconciler and the hook RPC alike — and proves it works before serving.
//
// sql.Open alone connects lazily, so a wrong DSN would first surface as a
// rejected push. Fail-closed makes that safe but not pleasant; failing at
// startup names the problem while somebody is still watching.
func openDatabase(dsn string) (*sql.DB, error) {
	pool, err := sql.Open("postgres", dsn)
	if err != nil {
		return nil, fmt.Errorf("open the database: %w", err)
	}
	ctx, cancel := context.WithTimeout(context.Background(), pingTimeout)
	defer cancel()
	if err := pool.PingContext(ctx); err != nil {
		pool.Close()
		return nil, fmt.Errorf("reach the database: %w", err)
	}
	return pool, nil
}

// refreshHooks installs this binary's hooks into every space.
//
// It runs at startup rather than only at space creation so that an upgrade
// which changes the wire protocol, the socket path or the hook set repairs
// every repository by restarting — there is no separate migration step and no
// repository left speaking last week's protocol.
func refreshHooks(ctx context.Context, log *slog.Logger, svc *service.Service, binary string) error {
	spaces, err := svc.ListSpaces(ctx)
	if err != nil {
		return fmt.Errorf("list spaces to refresh their hooks: %w", err)
	}
	for _, sp := range spaces {
		if err := hooks.InstallSpace(svc.ReposRoot(), sp.Ref, hooks.InstallOptions{Binary: binary}); err != nil {
			return fmt.Errorf("install the receive hooks of %s: %w "+
				"(a space whose hooks are missing would accept unvalidated pushes, so this is fatal; "+
				"repair or remove the repository and start again)", sp.Ref, err)
		}
	}
	log.Info("receive hooks refreshed", "spaces", len(spaces), "binary", binary)
	return nil
}

// mountRoutes installs what the daemon serves over HTTP.
func mountRoutes(router chi.Router, conf ini.File, surfaces *surfaces) {
	// server.New already froze the anonymous router for direct middleware
	// registration, so middleware and routes go in together inside a Group —
	// which chi permits on a fresh inline mux sharing the same routing tree.
	router.Group(func(r chi.Router) {
		r.Use(chimw.RealIP)
		r.Use(chimw.Recoverer)
		r.Get("/healthz", func(w http.ResponseWriter, _ *http.Request) {
			w.Header().Set("Content-Type", "text/plain; charset=utf-8")
			fmt.Fprint(w, "ok")
		})
	})

	mountWeb(router, conf, surfaces)
}

// surfaces are the three Phase 2 read surfaces, assembled once at startup.
//
// They share one *search.Index deliberately: bleve is single-writer, so a second
// Open on the same directory is not merely wasteful but wrong.
type surfaces struct {
	index *search.Index
	web   *web.Server
	mcp   http.Handler
	gql   http.Handler
}

// newSurfaces opens the index and builds the three read surfaces over it.
//
// All three go through service/ and none of them re-derives addressing, the
// read contract or the project filter — that shared layer is the whole reason
// "read SPEC-0007" cannot mean three different things depending on which door
// you knock on.
func newSurfaces(conf ini.File, cfg service.Config, svc *service.Service, version string) (*surfaces, error) {
	indexPath := filepath.Join(cfg.Cache, "index")
	// A crashed rebuild leaves working directories beside the index; clearing
	// them before Open is always safe, since the index is a pure cache.
	if err := search.CleanStale(indexPath); err != nil {
		return nil, fmt.Errorf("clean stale index working dirs: %w", err)
	}
	index, err := search.Open(indexPath)
	if err != nil {
		return nil, fmt.Errorf("open search index at %s: %w", indexPath, err)
	}

	site, err := web.New(web.Options{
		Conf:     conf,
		Reader:   web.NewReader(svc),
		Searcher: index,
		Resolver: svc.Resolver(),
	})
	if err != nil {
		index.Close()
		return nil, fmt.Errorf("assemble the web UI: %w", err)
	}

	// The origin is the /mcp Host allowlist: the MCP SDK's own DNS-rebinding
	// guard cannot tell a reverse proxy from an attacker (both reach a loopback
	// listener with a non-loopback Host), so it is replaced by a check against
	// this value. Traefik must pass the Host header through or every call 403s.
	mcp, err := mcpsrv.Handler(mcpsrv.Backend{Docs: svc, Index: index}, version, cfg.Origin)
	if err != nil {
		index.Close()
		return nil, fmt.Errorf("assemble the MCP surface: %w", err)
	}

	gql, err := graph.New(graph.Options{
		Reader:   svc,
		Searcher: index,
		Resolver: svc.Resolver(),
	})
	if err != nil {
		index.Close()
		return nil, fmt.Errorf("assemble the GraphQL surface: %w", err)
	}

	return &surfaces{index: index, web: site, mcp: mcp, gql: gql.Handler()}, nil
}

func (s *surfaces) Close() error {
	if s == nil || s.index == nil {
		return nil
	}
	return s.index.Close()
}

// mountWeb attaches the three read surfaces.
//
// Order is load-bearing: /mcp and /query are registered before the web UI,
// which mounts at "/" and would otherwise swallow them as document paths —
// "spec" and "query" are legal space names as far as the router is concerned.
//
// Each surface installs its own authentication (they are fail-closed and agree
// on one ACL), so core-go's WithDefaultMiddleware is deliberately not used: its
// auth middleware 401s any un-cookied request, which would also block the agent
// bearer-token path these surfaces exist to serve.
func mountWeb(router chi.Router, _ ini.File, s *surfaces) {
	if s == nil {
		return
	}
	router.Handle("/mcp", s.mcp)
	router.Handle("/query", s.gql)
	router.Mount("/", s.web.Handler())
}

// pushNotifier is what the daemon does when a push lands.
//
// Phase 1 records it and nothing more. Reindexing and advancing the space's
// index rev stamp are Phase 2's, because bleve and the stamp arrive together:
// moving the stamp now, with no index behind it, would assert that the index
// is current and remove the reconciler's only way of noticing that it is not.
func pushNotifier(log *slog.Logger) hooks.PushNotifier {
	return func(_ context.Context, space core.SpaceRef, updates []hooks.RefUpdate) error {
		refs := make([]string, 0, len(updates))
		for _, u := range updates {
			refs = append(refs, u.String())
		}
		log.Info("push landed; reindex pending",
			"space", space.String(),
			"refs", refs,
			"note", "indexing lands in Phase 2; the reconciler reports the staleness until then")
		return nil
	}
}

// reconcileReporter logs the outcome of each reconciler pass. A failure is not
// fatal: the next pass tries again, and a daemon that exits on a transient
// Postgres error takes the push path down with it.
func reconcileReporter(log *slog.Logger) func(*service.ReconcileReport, error) {
	return func(rep *service.ReconcileReport, err error) {
		if err != nil {
			log.Error("reconciler pass failed", scribe.Err(err))
			return
		}
		attrs := []any{
			"spaces", rep.Spaces,
			"repaired", len(rep.Repaired),
			"stale_indexes", len(rep.Reindex),
			"failures", len(rep.Failures),
		}
		for _, r := range rep.Repaired {
			log.Info("reconciler repaired divergence", "repair", fmt.Sprint(r))
		}
		for _, f := range rep.Failures {
			log.Warn("reconciler could not repair", "failure", fmt.Sprint(f))
		}
		log.Info("reconciler pass complete", attrs...)
	}
}

// bridgeSIGTERM makes systemd's default stop signal work.
//
// core-go's server.Run listens for SIGINT only, which is why compare.sr.ht's
// unit carries KillSignal=SIGINT. Rather than require that line here, a
// SIGTERM is turned into the SIGINT server.Run is waiting for. Both signals
// are registered before Run installs its own handler, so a signal arriving in
// the gap is caught rather than killing the process outright.
//
// After the first shutdown signal server.Run calls signal.Reset(os.Interrupt),
// restoring the default disposition — so a second signal terminates
// immediately, which is the documented behaviour and is why this keeps
// forwarding rather than stopping after one.
func bridgeSIGTERM(log *slog.Logger) {
	sig := make(chan os.Signal, 2)
	signal.Notify(sig, syscall.SIGTERM, os.Interrupt)
	go func() {
		for s := range sig {
			if s != syscall.SIGTERM {
				continue
			}
			log.Info("SIGTERM received; starting the warm shutdown core-go waits for SIGINT to begin")
			if err := syscall.Kill(os.Getpid(), syscall.SIGINT); err != nil {
				log.Error("could not raise SIGINT for the warm shutdown", scribe.Err(err))
			}
		}
	}()
}