ZDDC/zddc/cmd/zddc-server/main.go

package main

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

	"codeberg.org/VARASYS/ZDDC/zddc/internal/apps"
	"codeberg.org/VARASYS/ZDDC/zddc/internal/archive"
	"codeberg.org/VARASYS/ZDDC/zddc/internal/config"
	"codeberg.org/VARASYS/ZDDC/zddc/internal/handler"
	"codeberg.org/VARASYS/ZDDC/zddc/internal/policy"
	"codeberg.org/VARASYS/ZDDC/zddc/internal/tlsutil"
	"codeberg.org/VARASYS/ZDDC/zddc/internal/zddc"

	"github.com/klauspost/compress/gzhttp"
	"gopkg.in/natefinch/lumberjack.v2"
)

// version is the binary's own version, injected at build time via
// `-ldflags="-X main.version=..."`. Defaults to "dev" for unreleased
// builds; release pipelines pass the result of `git describe --tags`.
var version = "dev"

func main() {
	// --print-rego: dump a bundled reference Rego policy and exit.
	// Cheap escape hatch for operators standing up an external OPA who want
	// a parity-tested baseline as a starting point for customization.
	//
	//   --print-rego               → standard cascade (commercial default)
	//   --print-rego=standard      → same
	//   --print-rego=federal       → parent-deny-is-absolute (NIST AC-6)
	for _, a := range os.Args[1:] {
		switch a {
		case "--print-rego", "--print-rego=standard":
			fmt.Print(policy.ReferenceRego)
			return
		case "--print-rego=federal":
			fmt.Print(policy.FederalRego)
			return
		}
	}

	cfg, err := config.Load(os.Args[1:])
	if errors.Is(err, config.ErrHelpRequested) {
		config.Usage(os.Stderr)
		os.Exit(0)
	}
	if errors.Is(err, config.ErrVersionRequested) {
		printVersions(os.Stdout)
		os.Exit(0)
	}
	if err != nil {
		fmt.Fprintf(os.Stderr, "configuration error: %v\n\nRun with --help for usage.\n", err)
		os.Exit(1)
	}

	logRing := setupLogger(cfg.LogLevel)

	embedded := apps.EmbeddedVersions()
	slog.Info("zddc-server starting",
		"version", version,
		"root", cfg.Root,
		"addr", cfg.Addr,
		"embedded_apps", embeddedVersionsForLog(embedded))

	// Build archive index
	slog.Info("building archive index...")
	start := time.Now()
	idx, err := archive.BuildIndex(cfg.Root)
	if err != nil {
		slog.Error("failed to build archive index", "err", err)
		os.Exit(1)
	}
	slog.Info("archive index built", "duration", time.Since(start))

	// Apps fetch+cache subsystem.
	appsServer, err := setupApps(cfg)
	if err != nil {
		slog.Error("failed to set up apps subsystem", "err", err)
		os.Exit(1)
	}

	// TLS config
	tlsCfg, useTLS, err := tlsutil.TLSConfig(cfg)
	if err != nil {
		slog.Error("failed to configure TLS", "err", err)
		os.Exit(1)
	}

	// Context for graceful shutdown
	ctx, cancel := signal.NotifyContext(context.Background(), syscall.SIGTERM, syscall.SIGINT)
	defer cancel()

	// Start file-system watcher
	watcher, err := archive.NewWatcher(cfg.Root, idx)
	if err != nil {
		slog.Warn("failed to start filesystem watcher (index will not auto-update)", "err", err)
	} else {
		go func() {
			if err := watcher.Start(ctx); err != nil && ctx.Err() == nil {
				slog.Error("watcher error", "err", err)
			}
		}()
	}


	// HTTP handler
	mux := http.NewServeMux()
	// Middleware chain (outermost → innermost):
	//   ACLMiddleware     — extract email from cfg.EmailHeader, store in
	//                       request context. Outermost so the email is
	//                       available to AccessLogMiddleware (Go's context
	//                       propagates DOWN the chain via r.WithContext, not
	//                       UP — so AccessLog can't read a context value set
	//                       by an inner middleware after next.ServeHTTP
	//                       returns).
	//   AccessLogMiddleware — structured per-request log; reads email from
	//                       the context the outer ACL middleware set.
	//   CORSMiddleware    — Origin / preflight handling.
	//   dispatch          — the actual request handler.
	auditLogger := setupAccessAuditLog(cfg.AccessLog)

	// Construct the policy decider once at startup. ZDDC_OPA_URL=internal
	// (default) routes decisions through the in-process Go evaluator;
	// http(s):// or unix:// values send each decision to an external
	// OPA-compatible server (federal customers, custom Rego policies).
	deciderCfg := policy.Config{
		URL:         cfg.OPAURL,
		FailOpen:    cfg.OPAFailOpen,
		CacheTTL:    cfg.OPACacheTTL,
		CascadeMode: cfg.CascadeMode,
	}
	// Translate "0" (operator opt-out) to "disable cache" (negative TTL is
	// the policy package's sentinel for "skip the wrapper").
	if deciderCfg.CacheTTL == 0 {
		deciderCfg.CacheTTL = -1
	}
	decider, err := policy.New(deciderCfg)
	if err != nil {
		slog.Error("invalid OPA URL", "url", cfg.OPAURL, "err", err)
		os.Exit(1)
	}
	slog.Info("policy decider ready",
		"mode", policyModeLabel(cfg.OPAURL),
		"url", cfg.OPAURL,
		"cache_ttl", cfg.OPACacheTTL,
		"cascade_mode", cfg.CascadeMode)

	// Innermost handler: dispatch.
	var inner http.Handler = http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		dispatch(cfg, idx, logRing, appsServer, w, r)
	})
	inner = handler.CORSMiddleware(cfg, inner)
	// HSTS only when zddc-server itself is the TLS-terminating endpoint.
	// Behind an upstream proxy terminating TLS (cfg.TLSMode=="none"), the
	// proxy is responsible for HSTS — adding it here would conflict.
	if useTLS {
		inner = handler.HSTSMiddleware(inner)
	}
	inner = handler.AccessLogMiddleware(auditLogger, inner)
	inner = handler.ACLMiddleware(cfg, decider, inner)
	mux.Handle("/", inner)

	gzWrapper, err := newGzipWrapper()
	if err != nil {
		slog.Error("gzhttp wrapper init", "err", err)
		os.Exit(1)
	}

	srv := &http.Server{
		Addr:      cfg.Addr,
		Handler:   gzWrapper(mux),
		TLSConfig: tlsCfg,
		// Conservative timeouts. ReadHeaderTimeout caps how long a slow
		// client can hold the connection before sending request headers
		// (the slowloris vector). Read/Write timeouts cap full-request
		// processing — directory listings + tool HTML serving complete
		// in milliseconds even with gzip, so 60s is generous. IdleTimeout
		// is the keep-alive ceiling between requests on the same conn.
		ReadHeaderTimeout: 10 * time.Second,
		ReadTimeout:       60 * time.Second,
		WriteTimeout:      60 * time.Second,
		IdleTimeout:       120 * time.Second,
	}

	// Serve in goroutine
	if useTLS {
		go func() {
			slog.Info("listening", "addr", cfg.Addr, "tls", true)
			if err := srv.ListenAndServeTLS("", ""); err != nil && err != http.ErrServerClosed {
				slog.Error("server error", "err", err)
				cancel()
			}
		}()
	} else {
		go func() {
			slog.Info("listening", "addr", cfg.Addr, "tls", false)
			if err := srv.ListenAndServe(); err != nil && err != http.ErrServerClosed {
				slog.Error("server error", "err", err)
				cancel()
			}
		}()
	}

	<-ctx.Done()
	slog.Info("shutting down...")

	shutdownCtx, shutdownCancel := context.WithTimeout(context.Background(), 30*time.Second)
	defer shutdownCancel()
	if err := srv.Shutdown(shutdownCtx); err != nil {
		slog.Error("shutdown error", "err", err)
	}
	slog.Info("stopped")
}

// setupAccessAuditLog constructs a slog.Logger writing JSON lines to a
// size-rotated file at the operator-configured path. Returns nil if no
// path is configured (operator opted out via --access-log=) —
// AccessLogMiddleware then logs only to stderr.
//
// Auto-creates the parent directory (mode 0750) if missing, so the
// default path of <ZDDC_ROOT>/.zddc.d/logs/access-<host>.log "just
// works" on a fresh deployment without operator setup.
//
// Every record is tagged with `host` (os.Hostname). When multiple
// zddc-server replicas serve the same dataset (and write to the same
// .zddc.d/logs/ directory via per-host filenames), the host field also
// makes downstream-aggregated streams disambiguable.
//
// Rotation: lumberjack — 100 MB per file, 10 backups, 90-day max age,
// gzip compression on rotated files.
//
// File-permission posture: lumberjack creates new logs with mode 0600
// (running user only). For multi-user audit access, the operator should
// use group-readable parent directory permissions and either chmod the
// log out-of-band or run a forwarder that has its own read access.
// policyModeLabel collapses cfg.OPAURL to a one-word mode label for the
// startup log so operators can grep for the active decider quickly.
func policyModeLabel(opaURL string) string {
	switch {
	case opaURL == "" || strings.EqualFold(opaURL, "internal"):
		return "internal"
	case strings.HasPrefix(opaURL, "unix://"):
		return "external-unix"
	case strings.HasPrefix(opaURL, "https://"):
		return "external-https"
	case strings.HasPrefix(opaURL, "http://"):
		return "external-http"
	default:
		return "unknown"
	}
}

func setupAccessAuditLog(path string) *slog.Logger {
	if path == "" {
		return nil
	}
	if err := os.MkdirAll(filepath.Dir(path), 0o750); err != nil {
		slog.Error("could not create access-log directory; falling back to stderr-only",
			"dir", filepath.Dir(path), "err", err)
		return nil
	}
	rotator := &lumberjack.Logger{
		Filename:   path,
		MaxSize:    100, // megabytes per file before rotation
		MaxBackups: 10,
		MaxAge:     90, // days
		Compress:   true,
	}
	host, _ := os.Hostname()
	if host == "" {
		host = "unknown"
	}
	// JSON handler — line-delimited JSON is the format every standard
	// log shipper (Vector, Loki promtail, fluentbit, journalbeat) parses
	// natively, and stays grep-friendly for ad-hoc inspection.
	h := slog.NewJSONHandler(rotator, &slog.HandlerOptions{Level: slog.LevelInfo})
	slog.Info("access log file enabled",
		"path", path, "host", host,
		"max_size_mb", 100, "max_backups", 10, "max_age_days", 90)
	return slog.New(h).With("host", host)
}

// newGzipWrapper builds the gzip middleware applied to the entire mux.
// MinSize(1024) skips compressing tiny responses where the framing
// overhead exceeds the savings (304 Not Modified, error pages, small
// JSON listings under ~1 KB). The wrapper honors Accept-Encoding (passes
// through unchanged when the client doesn't advertise gzip), appends
// Vary: Accept-Encoding automatically, and passes through 304s untouched.
// Yields ~75% size reduction on the larger embedded HTML responses
// (mdedit: 920 KB → ~250 KB on the wire).
//
// Extracted so tests can construct an equivalent wrapper without going
// through the full main() server boot.
func newGzipWrapper() (func(http.Handler) http.HandlerFunc, error) {
	return gzhttp.NewWrapper(gzhttp.MinSize(1024))
}

// setupApps creates the cache + fetcher + server. No seeding, no refresh,
// no admin UI — the server fetches once on first request, caches forever
// in <ZDDC_ROOT>/_app/, and falls back to the embedded HTML on any failure.
func setupApps(cfg config.Config) (*apps.Server, error) {
	cache, err := apps.NewCache(filepath.Join(cfg.Root, apps.CacheDirName))
	if err != nil {
		return nil, fmt.Errorf("create cache: %w", err)
	}
	fetcher := apps.NewFetcher(cache, slog.Default())

	// Apps signing pubkey. Resolution order, highest priority first:
	//   1. --apps-pubkey / ZDDC_APPS_PUBKEY (path to PEM file)
	//   2. apps_pubkey: inline PEM in the root <ZDDC_ROOT>/.zddc file
	//      (root-only — same trust-anchor treatment as admins:)
	//   3. nothing → URL-fetched apps refuse-by-default; only embedded
	//      + local-path apps work
	//
	// Same posture as TLS certificates: zddc-server bakes nothing in.
	// Operators using zddc.varasys.io's canonical channels download
	// pubkey.pem from there and either configure the path via env/flag
	// or paste the PEM contents inline into root .zddc.
	switch {
	case cfg.AppsPubKey != "":
		pub, err := apps.LoadPubKey(cfg.AppsPubKey)
		if err != nil {
			return nil, fmt.Errorf("apps-pubkey: %w", err)
		}
		fetcher.VerifyKey = pub
		slog.Info("apps signing pubkey loaded", "source", "env/flag", "path", cfg.AppsPubKey)
	default:
		// Fall back to apps_pubkey: in root .zddc.
		rootZddc, err := zddc.ParseFile(filepath.Join(cfg.Root, ".zddc"))
		if err == nil && rootZddc.AppsPubKey != "" {
			pub, err := apps.ParsePubKeyPEM([]byte(rootZddc.AppsPubKey))
			if err != nil {
				return nil, fmt.Errorf("root .zddc apps_pubkey: %w", err)
			}
			fetcher.VerifyKey = pub
			slog.Info("apps signing pubkey loaded", "source", "root .zddc apps_pubkey")
		} else {
			slog.Warn("apps-pubkey not configured; URL-fetched apps will be refused (only embedded + local-path apps will work). " +
				"Set --apps-pubkey, ZDDC_APPS_PUBKEY, or apps_pubkey: in the root .zddc file to a PEM Ed25519 public key you trust.")
		}
	}

	return apps.NewServer(cfg.Root, cache, fetcher, version), nil
}

// printVersions writes the binary version + the build label of every app
// embedded into the binary. Called by --version and reused for the
// startup log line.
func printVersions(w *os.File) {
	fmt.Fprintf(w, "zddc-server %s\n\n", version)
	embedded := apps.EmbeddedVersions()
	if len(embedded) == 0 {
		fmt.Fprintln(w, "Embedded tools: (none — run `sh build.sh` to populate)")
		return
	}
	fmt.Fprintln(w, "Embedded tools:")
	keys := make([]string, 0, len(embedded))
	for k := range embedded {
		keys = append(keys, k)
	}
	sort.Strings(keys)
	for _, k := range keys {
		fmt.Fprintf(w, "  %-12s %s\n", k, embedded[k])
	}
}

// embeddedVersionsForLog formats the embedded-versions map as a single
// short string suitable for the startup `log/slog` line. Sorted by app
// name for stable output.
func embeddedVersionsForLog(embedded map[string]string) string {
	if len(embedded) == 0 {
		return "(none)"
	}
	keys := make([]string, 0, len(embedded))
	for k := range embedded {
		keys = append(keys, k)
	}
	sort.Strings(keys)
	parts := make([]string, 0, len(keys))
	for _, k := range keys {
		// Strip any " · timestamp · sha" suffix so the log line stays compact;
		// operators who want full detail run `zddc-server --version`.
		v := embedded[k]
		if i := strings.Index(v, " "); i > 0 {
			v = v[:i]
		}
		parts = append(parts, k+"="+v)
	}
	return strings.Join(parts, " ")
}

// dispatch routes a request to the appropriate handler.
func dispatch(cfg config.Config, idx *archive.Index, ring *handler.LogRing, appsSrv *apps.Server, w http.ResponseWriter, r *http.Request) {
	urlPath := r.URL.Path
	email := handler.EmailFromContext(r)

	// Profile routes — the page itself is reachable to anyone (anonymous
	// included); admin-only sub-resources (whoami / config / logs /
	// projects / .zddc editor) keep their existing per-resource 404
	// existence-leakage gates inside ServeProfile.
	if urlPath == handler.ProfilePathPrefix || strings.HasPrefix(urlPath, handler.ProfilePathPrefix+"/") {
		handler.ServeProfile(cfg, ring, w, r)
		return
	}

	// Auth check endpoints — machine-only forward_auth targets used by
	// upstream proxies (e.g. the dev-shell pod's Caddy in front of
	// code-server) to gate routes on root-admin status. Handled before
	// the reserved-prefix guard below so the .auth namespace passes
	// through without being 404'd by the dot-prefix rule.
	if urlPath == handler.AuthPathPrefix+"/admin" {
		handler.ServeAuthAdmin(cfg, w, r)
		return
	}

	// Project list API: GET / with Accept: application/json
	if urlPath == "/" {
		accept := r.Header.Get("Accept")
		if strings.Contains(accept, "application/json") {
			handler.ServeProjectList(cfg, w, r)
			return
		}
	}

	// Split path into segments
	segments := strings.Split(strings.Trim(urlPath, "/"), "/")

	// Reserve dot-prefixed path segments. The listing pipeline already hides
	// hidden entries (internal/listing/listing.go:17, projectshandler.go:40),
	// but direct URL access would still serve them. 404 here so hidden trees
	// like /srv/.devshell (the in-image dev-shell's persistent home dir on
	// the same Azure Files PVC as served data) cannot be fetched. The
	// recognized virtual prefixes (.profile handled above, cfg.IndexPath
	// handled below) are explicitly allowed through.
	//
	// Also reserve the apps cache directory (`_app`): the cached HTML files
	// there must be served via the apps resolver (with proper headers and
	// ACL), never raw at /_app/...html.
	for _, seg := range segments {
		if seg == "" {
			continue
		}
		if seg == apps.CacheDirName {
			http.NotFound(w, r)
			return
		}
		if !strings.HasPrefix(seg, ".") {
			continue
		}
		if seg == cfg.IndexPath {
			continue
		}
		http.NotFound(w, r)
		return
	}

	// Check for .archive segment in the path
	for i, seg := range segments {
		if seg == cfg.IndexPath {
			// contextPath is everything before .archive
			contextPath := "/" + strings.Join(segments[:i], "/")
			var filename string
			if i+1 < len(segments) {
				filename = strings.Join(segments[i+1:], "/")
			}
			handler.ServeArchive(cfg, idx, w, r, contextPath, filename)
			return
		}
	}

	// Tables-system intercept: *.table.html is a virtual URL that the
	// table handler renders inline, reading rows from a directory of
	// *.yaml files declared in the directory's .zddc tables: map.
	// Discovery is .zddc-declarative — no auto-mount on file presence —
	// so RecognizeTableRequest returns nil whenever there's no matching
	// declaration and the URL falls through to the static-file path
	// (or to the form intercept below for *.form.html / *.yaml.html).
	if tableReq := handler.RecognizeTableRequest(cfg.Root, r.Method, urlPath); tableReq != nil {
		handler.ServeTable(cfg, tableReq, w, r)
		return
	}

	// Form-system intercept: *.form.html and *.yaml.html under a sibling form
	// folder are virtual URLs that the form handler renders inline, reading
	// the underlying *.form.yaml spec (and, for re-edit, the *.yaml data) from
	// disk. RecognizeFormRequest returns nil when the spec doesn't exist, so
	// non-form .html URLs fall through to the static-file path below.
	if formReq := handler.RecognizeFormRequest(cfg.Root, r.Method, urlPath); formReq != nil {
		handler.ServeForm(cfg, formReq, w, r)
		return
	}

	// File API — authenticated CRUD over the served tree. Catches PUT,
	// DELETE, and POST on any non-reserved path. Read methods (GET/HEAD)
	// fall through to the static / apps / directory pipeline below.
	// Forms and .profile/.archive POSTs are already routed above this
	// point so they take precedence.
	if handler.IsWriteMethod(r.Method) {
		handler.ServeFileAPI(cfg, w, r)
		return
	}

	// Apps resolution for the root landing path: GET / or /index.html with
	// no real index.html on disk → serve via apps.Serve("landing"). The
	// other four apps are caught by the "stat fails → app HTML?" branch
	// below, which only triggers when no concrete file is at the URL path.
	//
	// The landing page is intentionally public (no ACL gate). It's a
	// project picker — the per-project ACL filtering done by
	// fs.ListDirectory still hides projects an anonymous (or unauthorized)
	// caller can't reach. See also handler.ServeDirectory's matching
	// root-path bypass.
	if appsSrv != nil && (urlPath == "/" || urlPath == "/index.html") {
		realIndex := filepath.Join(cfg.Root, "index.html")
		if _, err := os.Stat(realIndex); os.IsNotExist(err) {
			chain, _ := zddc.EffectivePolicy(cfg.Root, cfg.Root)
			if apps.AppAvailableAt(cfg.Root, cfg.Root, "landing") {
				appsSrv.Serve(w, r, "landing", chain, cfg.Root)
				return
			}
		}
	}

	// Resolve the physical path
	cleanPath := filepath.FromSlash(strings.TrimPrefix(urlPath, "/"))
	absPath := filepath.Join(cfg.Root, cleanPath)

	// Guard against path traversal
	if !strings.HasPrefix(absPath, cfg.Root+string(filepath.Separator)) && absPath != cfg.Root {
		http.Error(w, "Not Found", http.StatusNotFound)
		return
	}

	// Check filesystem
	info, err := os.Stat(absPath)
	if err != nil {
		if os.IsNotExist(err) {
			// File doesn't exist at this path. If the URL matches one of
			// the five canonical app HTML names AND the request directory
			// is one where that app is available (Incoming/Working/Staging
			// for classifier/mdedit/transmittal, anywhere for archive,
			// root only for landing), resolve via the apps subsystem.
			if appsSrv != nil {
				if app, requestDirRel := apps.MatchAppHTML(urlPath); app != "" {
					requestDir := filepath.Join(cfg.Root, filepath.FromSlash(requestDirRel))
					if apps.AppAvailableAt(cfg.Root, requestDir, app) {
						chain, _ := zddc.EffectivePolicy(cfg.Root, requestDir)
						if allowed, _ := policy.AllowFromChain(r.Context(), handler.DeciderFromContext(r), chain, email, urlPath); !allowed {
							http.Error(w, "Forbidden", http.StatusForbidden)
							return
						}
						appsSrv.Serve(w, r, app, chain, requestDir)
						return
					}
				}
			}
			http.Error(w, "Not Found", http.StatusNotFound)
		} else {
			http.Error(w, "Internal Server Error", http.StatusInternalServerError)
		}
		return
	}

	if info.IsDir() {
		// ACL check — bypassed at the root path so the landing page (the
		// project picker) is reachable by anyone, including anonymous.
		// Per-project filtering happens inside ServeDirectory →
		// fs.ListDirectory, which hides directories the caller can't
		// reach. Subdirectory requests still hit this gate.
		isRoot := urlPath == "/"
		if !isRoot {
			chain, _ := zddc.EffectivePolicy(cfg.Root, absPath)
			if allowed, _ := policy.AllowFromChain(r.Context(), handler.DeciderFromContext(r), chain, email, urlPath); !allowed {
				http.Error(w, "Forbidden", http.StatusForbidden)
				return
			}
		}
		if !strings.HasSuffix(urlPath, "/") {
			http.Redirect(w, r, urlPath+"/", http.StatusMovedPermanently)
			return
		}
		handler.ServeDirectory(cfg, w, r)
		return
	}

	// Regular file: ACL on parent directory
	chain, _ := zddc.EffectivePolicy(cfg.Root, filepath.Dir(absPath))
	if allowed, _ := policy.AllowFromChain(r.Context(), handler.DeciderFromContext(r), chain, email, urlPath); !allowed {
		http.Error(w, "Forbidden", http.StatusForbidden)
		return
	}
	handler.ServeFile(w, r, absPath)
}

// setupLogger installs a slog default that fans every record out to stderr
// (the existing TextHandler — user-visible logging is unchanged) AND to an
// in-memory ring buffer that backs the /.profile/logs endpoint. Returns
// the ring so handlers can read it.
func setupLogger(level string) *handler.LogRing {
	var l slog.Level
	switch strings.ToLower(level) {
	case "debug":
		l = slog.LevelDebug
	case "warn":
		l = slog.LevelWarn
	case "error":
		l = slog.LevelError
	default:
		l = slog.LevelInfo
	}
	ring := handler.NewLogRing(500)
	text := slog.NewTextHandler(os.Stderr, &slog.HandlerOptions{Level: l})
	rh := handler.NewRingHandler(ring, l)
	slog.SetDefault(slog.New(handler.NewMultiHandler(text, rh)))
	return ring
}