Add transparent proxy inspection engine with envoy sidecar support

client/inspect/config.go

@@ -0,0 +1,212 @@
+package inspect
+
+import (
+	"crypto"
+	"crypto/x509"
+	"net"
+	"net/netip"
+	"net/url"
+	"strings"
+
+	"github.com/netbirdio/netbird/client/internal/acl/id"
+	"github.com/netbirdio/netbird/shared/management/domain"
+)
+
+// InspectResult holds the outcome of connection inspection.
+type InspectResult struct {
+	// Action is the rule evaluation result.
+	Action Action
+	// PassthroughConn is the client connection with buffered peeked bytes.
+	// Non-nil only when Action is ActionAllow and the caller should relay
+	// (TLS passthrough or non-HTTP/TLS protocol). The caller takes ownership
+	// and is responsible for closing this connection.
+	PassthroughConn net.Conn
+}
+
+const (
+	// DefaultTProxyPort is the default TPROXY listener port for kernel mode.
+	// Override with NB_TPROXY_PORT environment variable.
+	DefaultTProxyPort = 22080
+)
+
+// Action determines how the proxy handles a matched connection.
+type Action string
+
+const (
+	// ActionAllow passes the connection through without decryption.
+	ActionAllow Action = "allow"
+	// ActionBlock denies the connection.
+	ActionBlock Action = "block"
+	// ActionInspect decrypts (MITM) and inspects the connection.
+	ActionInspect Action = "inspect"
+)
+
+// ProxyMode determines the proxy operating mode.
+type ProxyMode string
+
+const (
+	// ModeBuiltin uses the built-in proxy with rules and optional ICAP.
+	ModeBuiltin ProxyMode = "builtin"
+	// ModeEnvoy runs a local envoy sidecar for L7 processing.
+	// Go manages envoy lifecycle, config generation, and rule evaluation.
+	// USP path forwards via PROXY protocol v2; kernel path uses nftables redirect.
+	ModeEnvoy ProxyMode = "envoy"
+	// ModeExternal forwards all traffic to an external proxy.
+	ModeExternal ProxyMode = "external"
+)
+
+// PolicyID is the management policy identifier associated with a connection.
+type PolicyID []byte
+
+// MatchDomain reports whether target matches the pattern.
+// If pattern starts with "*.", it matches any subdomain (but not the base itself).
+// Otherwise it requires an exact match.
+func MatchDomain(pattern, target domain.Domain) bool {
+	p := pattern.PunycodeString()
+	t := target.PunycodeString()
+
+	if strings.HasPrefix(p, "*.") {
+		base := p[2:]
+		return strings.HasSuffix(t, "."+base)
+	}
+
+	return p == t
+}
+
+// SourceInfo carries source identity context for rule evaluation.
+// The source may be a direct WireGuard peer or a host behind
+// a site-to-site gateway.
+type SourceInfo struct {
+	// IP is the original source address from the packet.
+	IP netip.Addr
+	// PolicyID is the management policy that allowed this traffic
+	// through route ACLs.
+	PolicyID PolicyID
+}
+
+// ProtoType identifies a protocol handled by the proxy.
+type ProtoType string
+
+const (
+	ProtoHTTP      ProtoType = "http"
+	ProtoHTTPS     ProtoType = "https"
+	ProtoH2        ProtoType = "h2"
+	ProtoH3        ProtoType = "h3"
+	ProtoWebSocket ProtoType = "websocket"
+	ProtoOther     ProtoType = "other"
+)
+
+// Rule defines a proxy inspection/filtering rule.
+type Rule struct {
+	// ID uniquely identifies this rule.
+	ID id.RuleID
+	// Sources are the source CIDRs this rule applies to.
+	// Includes both direct peer IPs and routed networks behind gateways.
+	Sources []netip.Prefix
+	// Domains are the destination domain patterns to match (via SNI or Host header).
+	// Supports exact match ("example.com") and wildcard ("*.example.com").
+	Domains []domain.Domain
+	// Networks are the destination CIDRs to match.
+	Networks []netip.Prefix
+	// Ports are the destination ports to match. Empty means all ports.
+	Ports []uint16
+	// Protocols restricts which protocols this rule applies to.
+	// Empty means all protocols.
+	Protocols []ProtoType
+	// Paths are URL path patterns to match (HTTP only, requires inspect for HTTPS).
+	// Supports prefix ("/api/"), exact ("/login"), and wildcard ("/admin/*").
+	// Empty means all paths.
+	Paths []string
+	// Action determines what to do with matched connections.
+	Action Action
+	// Priority controls evaluation order. Lower values are evaluated first.
+	Priority int
+}
+
+// ICAPConfig holds ICAP service configuration.
+type ICAPConfig struct {
+	// ReqModURL is the ICAP REQMOD service URL (e.g., icap://server:1344/reqmod).
+	ReqModURL *url.URL
+	// RespModURL is the ICAP RESPMOD service URL (e.g., icap://server:1344/respmod).
+	RespModURL *url.URL
+	// MaxConnections is the connection pool size. Zero uses a default.
+	MaxConnections int
+}
+
+// TLSConfig holds the MITM CA configuration for TLS inspection.
+type TLSConfig struct {
+	// CA is the certificate authority used to sign dynamic certificates.
+	CA *x509.Certificate
+	// CAKey is the CA's private key.
+	CAKey crypto.PrivateKey
+}
+
+// Config holds the transparent proxy configuration.
+type Config struct {
+	// Enabled controls whether the proxy is active.
+	Enabled bool
+	// Mode selects built-in or external proxy operation.
+	Mode ProxyMode
+	// ExternalURL is the upstream proxy URL for ModeExternal.
+	// Supports http:// and socks5:// schemes.
+	ExternalURL *url.URL
+
+	// DefaultAction applies when no rule matches a connection.
+	DefaultAction Action
+
+	// RedirectSources are the source CIDRs whose traffic should be intercepted.
+	// Admin decides: "activate for these users/subnets."
+	// Used for both kernel TPROXY rules and userspace forwarder source filtering.
+	RedirectSources []netip.Prefix
+	// RedirectPorts are the destination ports to intercept. Empty means all ports.
+	RedirectPorts []uint16
+
+	// Rules are the proxy inspection/filtering rules, evaluated in Priority order.
+	Rules []Rule
+
+	// ICAP holds ICAP service configuration. Nil disables ICAP.
+	ICAP *ICAPConfig
+	// TLS holds the MITM CA. Nil means no MITM capability (ActionInspect rules ignored).
+	TLS *TLSConfig
+
+	// Envoy configuration (ModeEnvoy only)
+	Envoy *EnvoyConfig
+
+	// ListenAddr is the TPROXY listen address for kernel mode.
+	// Zero value disables the TPROXY listener.
+	ListenAddr netip.AddrPort
+	// WGNetwork is the WireGuard overlay network prefix.
+	// The proxy blocks dialing destinations inside this network.
+	WGNetwork netip.Prefix
+	// LocalIPChecker reports whether an IP belongs to the routing peer.
+	// Used to prevent SSRF to local services. May be nil.
+	LocalIPChecker LocalIPChecker
+}
+
+// EnvoyConfig holds configuration for the envoy sidecar mode.
+type EnvoyConfig struct {
+	// BinaryPath is the path to the envoy binary.
+	// Empty means search $PATH for "envoy".
+	BinaryPath string
+	// AdminPort is the port for envoy's admin API (health checks, stats).
+	// Zero means auto-assign.
+	AdminPort uint16
+	// Snippets are user-provided config fragments merged into the generated bootstrap.
+	Snippets *EnvoySnippets
+}
+
+// EnvoySnippets holds user-provided YAML fragments for envoy config customization.
+// Only safe snippet types are allowed: filters (HTTP and network) and clusters
+// needed as dependencies for filter services. Listeners and bootstrap overrides
+// are not exposed since we manage the listener and bootstrap.
+type EnvoySnippets struct {
+	// HTTPFilters is YAML injected into the HCM filter chain before the router filter.
+	// Used for ext_authz, rate limiting, Lua, Wasm, RBAC, JWT auth, etc.
+	HTTPFilters string
+	// NetworkFilters is YAML injected into the TLS filter chain before tcp_proxy.
+	// Used for network-level RBAC, rate limiting, ext_authz on raw TCP.
+	NetworkFilters string
+	// Clusters is YAML for additional upstream clusters referenced by filters.
+	// Needed when filters call external services (ext_authz backend, rate limit service).
+	Clusters string
+}

client/inspect/config_test.go

@@ -0,0 +1,93 @@
+package inspect
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/shared/management/domain"
+)
+
+func TestMatchDomain(t *testing.T) {
+	tests := []struct {
+		name    string
+		pattern string
+		target  string
+		want    bool
+	}{
+		{
+			name:    "exact match",
+			pattern: "example.com",
+			target:  "example.com",
+			want:    true,
+		},
+		{
+			name:    "exact no match",
+			pattern: "example.com",
+			target:  "other.com",
+			want:    false,
+		},
+		{
+			name:    "wildcard matches subdomain",
+			pattern: "*.example.com",
+			target:  "foo.example.com",
+			want:    true,
+		},
+		{
+			name:    "wildcard matches deep subdomain",
+			pattern: "*.example.com",
+			target:  "a.b.c.example.com",
+			want:    true,
+		},
+		{
+			name:    "wildcard does not match base",
+			pattern: "*.example.com",
+			target:  "example.com",
+			want:    false,
+		},
+		{
+			name:    "wildcard does not match unrelated",
+			pattern: "*.example.com",
+			target:  "foo.other.com",
+			want:    false,
+		},
+		{
+			name:    "case insensitive exact match",
+			pattern: "Example.COM",
+			target:  "example.com",
+			want:    true,
+		},
+		{
+			name:    "case insensitive wildcard match",
+			pattern: "*.Example.COM",
+			target:  "FOO.example.com",
+			want:    true,
+		},
+		{
+			name:    "wildcard does not match partial suffix",
+			pattern: "*.example.com",
+			target:  "notexample.com",
+			want:    false,
+		},
+		{
+			name:    "unicode domain punycode match",
+			pattern: "*.münchen.de",
+			target:  "sub.xn--mnchen-3ya.de",
+			want:    true,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			pattern, err := domain.FromString(tt.pattern)
+			require.NoError(t, err)
+
+			target, err := domain.FromString(tt.target)
+			require.NoError(t, err)
+
+			got := MatchDomain(pattern, target)
+			assert.Equal(t, tt.want, got)
+		})
+	}
+}

client/inspect/dialer_linux.go

@@ -0,0 +1,25 @@
+package inspect
+
+import (
+	"net"
+	"syscall"
+)
+
+// newOutboundDialer creates a net.Dialer that clears the socket fwmark.
+// In kernel TPROXY mode, accepted connections inherit the TPROXY fwmark.
+// Without clearing it, outbound connections from the proxy would match
+// the ip rule (fwmark -> local loopback) and loop back to the proxy
+// instead of reaching the real destination.
+func newOutboundDialer() net.Dialer {
+	return net.Dialer{
+		Control: func(_, _ string, c syscall.RawConn) error {
+			var sockErr error
+			if err := c.Control(func(fd uintptr) {
+				sockErr = syscall.SetsockoptInt(int(fd), syscall.SOL_SOCKET, syscall.SO_MARK, 0)
+			}); err != nil {
+				return err
+			}
+			return sockErr
+		},
+	}
+}

client/inspect/dialer_other.go

@@ -0,0 +1,11 @@
+//go:build !linux
+
+package inspect
+
+import "net"
+
+// newOutboundDialer returns a plain dialer on non-Linux platforms.
+// TPROXY is Linux-only, so no fwmark clearing is needed.
+func newOutboundDialer() net.Dialer {
+	return net.Dialer{}
+}

client/inspect/envoy.go

@@ -0,0 +1,298 @@
+package inspect
+
+import (
+	"context"
+	"fmt"
+	"io"
+	"net"
+	"net/http"
+	"net/netip"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"sync"
+	"time"
+
+	log "github.com/sirupsen/logrus"
+)
+
+const (
+	envoyStartTimeout   = 15 * time.Second
+	envoyHealthInterval = 500 * time.Millisecond
+	envoyStopTimeout    = 10 * time.Second
+	envoyDrainTime      = 5
+)
+
+// envoyManager manages the lifecycle of an envoy sidecar process.
+type envoyManager struct {
+	log        *log.Entry
+	cmd        *exec.Cmd
+	configPath string
+	listenPort uint16
+	adminPort  uint16
+	cancel     context.CancelFunc
+
+	blockPagePath string
+
+	mu      sync.Mutex
+	running bool
+}
+
+// startEnvoy finds the envoy binary, generates config, and spawns the process.
+// It blocks until envoy reports healthy or the timeout expires.
+func startEnvoy(ctx context.Context, logger *log.Entry, config Config) (*envoyManager, error) {
+	envCfg := config.Envoy
+	if envCfg == nil {
+		return nil, fmt.Errorf("envoy config is nil")
+	}
+
+	binaryPath, err := findEnvoyBinary(envCfg.BinaryPath)
+	if err != nil {
+		return nil, fmt.Errorf("find envoy binary: %w", err)
+	}
+
+	// Pick admin port
+	adminPort := envCfg.AdminPort
+	if adminPort == 0 {
+		p, err := findFreePort()
+		if err != nil {
+			return nil, fmt.Errorf("find free admin port: %w", err)
+		}
+		adminPort = p
+	}
+
+	// Pick listener port
+	listenPort, err := findFreePort()
+	if err != nil {
+		return nil, fmt.Errorf("find free listener port: %w", err)
+	}
+
+	// Use a private temp directory (0700) to prevent local attackers from
+	// replacing the config file between write and envoy read.
+	configDir, err := os.MkdirTemp("", "nb-envoy-*")
+	if err != nil {
+		return nil, fmt.Errorf("create envoy config directory: %w", err)
+	}
+
+	// Write the block page HTML for envoy's direct_response to reference.
+	blockPagePath := filepath.Join(configDir, "block.html")
+	blockHTML := fmt.Sprintf(blockPageHTML, "blocked domain", "this domain")
+	if err := os.WriteFile(blockPagePath, []byte(blockHTML), 0600); err != nil {
+		return nil, fmt.Errorf("write envoy block page: %w", err)
+	}
+
+	// Generate config with the block page path embedded.
+	bootstrap, err := generateBootstrap(config, listenPort, adminPort, blockPagePath)
+	if err != nil {
+		return nil, fmt.Errorf("generate envoy bootstrap: %w", err)
+	}
+
+	configPath := filepath.Join(configDir, "bootstrap.yaml")
+	if err := os.WriteFile(configPath, bootstrap, 0600); err != nil {
+		return nil, fmt.Errorf("write envoy config: %w", err)
+	}
+
+	ctx, cancel := context.WithCancel(ctx)
+
+	cmd := exec.CommandContext(ctx, binaryPath,
+		"-c", configPath,
+		"--drain-time-s", fmt.Sprintf("%d", envoyDrainTime),
+	)
+
+	// Pipe envoy output to our logger.
+	cmd.Stdout = &logWriter{entry: logger, level: log.DebugLevel}
+	cmd.Stderr = &logWriter{entry: logger, level: log.WarnLevel}
+
+	if err := cmd.Start(); err != nil {
+		cancel()
+		os.Remove(configPath)
+		return nil, fmt.Errorf("start envoy: %w", err)
+	}
+
+	mgr := &envoyManager{
+		log:           logger,
+		cmd:           cmd,
+		configPath:    configPath,
+		listenPort:    listenPort,
+		adminPort:     adminPort,
+		blockPagePath: blockPagePath,
+		cancel:        cancel,
+		running:       true,
+	}
+
+	// Wait for envoy to become healthy.
+	if err := mgr.waitHealthy(ctx); err != nil {
+		mgr.Stop()
+		return nil, fmt.Errorf("wait for envoy readiness: %w", err)
+	}
+
+	logger.Infof("inspect: envoy started (pid=%d, listen=%d, admin=%d)", cmd.Process.Pid, listenPort, adminPort)
+
+	// Monitor process exit in background.
+	go mgr.monitor()
+
+	return mgr, nil
+}
+
+// ListenAddr returns the address envoy listens on for forwarded connections.
+func (m *envoyManager) ListenAddr() netip.AddrPort {
+	return netip.AddrPortFrom(netip.AddrFrom4([4]byte{127, 0, 0, 1}), m.listenPort)
+}
+
+// AdminAddr returns the envoy admin API address.
+func (m *envoyManager) AdminAddr() string {
+	return fmt.Sprintf("127.0.0.1:%d", m.adminPort)
+}
+
+// Reload writes a new config and sends SIGHUP to envoy.
+func (m *envoyManager) Reload(config Config) error {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+
+	if !m.running {
+		return fmt.Errorf("envoy is not running")
+	}
+
+	bootstrap, err := generateBootstrap(config, m.listenPort, m.adminPort, m.blockPagePath)
+	if err != nil {
+		return fmt.Errorf("generate envoy bootstrap: %w", err)
+	}
+
+	if err := os.WriteFile(m.configPath, bootstrap, 0600); err != nil {
+		return fmt.Errorf("write envoy config: %w", err)
+	}
+
+	if err := signalReload(m.cmd.Process); err != nil {
+		return fmt.Errorf("signal envoy reload: %w", err)
+	}
+
+	m.log.Debugf("inspect: envoy config reloaded")
+	return nil
+}
+
+// Healthy checks the envoy admin API /ready endpoint.
+func (m *envoyManager) Healthy() bool {
+	resp, err := http.Get(fmt.Sprintf("http://%s/ready", m.AdminAddr()))
+	if err != nil {
+		return false
+	}
+	defer resp.Body.Close()
+	return resp.StatusCode == http.StatusOK
+}
+
+// Stop terminates the envoy process and cleans up.
+func (m *envoyManager) Stop() {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+
+	if !m.running {
+		return
+	}
+	m.running = false
+
+	m.cancel()
+
+	if m.cmd.Process != nil {
+		done := make(chan struct{})
+		go func() {
+			m.cmd.Wait()
+			close(done)
+		}()
+
+		select {
+		case <-done:
+		case <-time.After(envoyStopTimeout):
+			m.log.Warnf("inspect: envoy did not exit in %s, killing", envoyStopTimeout)
+			m.cmd.Process.Kill()
+			<-done
+		}
+	}
+
+	os.RemoveAll(filepath.Dir(m.configPath))
+	m.log.Infof("inspect: envoy stopped")
+}
+
+// waitHealthy polls the admin API until envoy is ready or timeout.
+func (m *envoyManager) waitHealthy(ctx context.Context) error {
+	deadline := time.After(envoyStartTimeout)
+	ticker := time.NewTicker(envoyHealthInterval)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case <-deadline:
+			return fmt.Errorf("envoy not ready after %s", envoyStartTimeout)
+		case <-ticker.C:
+			if m.Healthy() {
+				return nil
+			}
+		}
+	}
+}
+
+// monitor watches for unexpected envoy exits.
+func (m *envoyManager) monitor() {
+	err := m.cmd.Wait()
+
+	m.mu.Lock()
+	wasRunning := m.running
+	m.running = false
+	m.mu.Unlock()
+
+	if wasRunning {
+		m.log.Errorf("inspect: envoy exited unexpectedly: %v", err)
+	}
+}
+
+// findEnvoyBinary resolves the envoy binary path.
+func findEnvoyBinary(configPath string) (string, error) {
+	if configPath != "" {
+		if _, err := os.Stat(configPath); err != nil {
+			return "", fmt.Errorf("envoy binary not found at %s: %w", configPath, err)
+		}
+		return configPath, nil
+	}
+
+	path, err := exec.LookPath("envoy")
+	if err != nil {
+		return "", fmt.Errorf("envoy not found in PATH: %w", err)
+	}
+	return path, nil
+}
+
+// findFreePort asks the OS for an available TCP port.
+func findFreePort() (uint16, error) {
+	ln, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		return 0, err
+	}
+	port := uint16(ln.Addr().(*net.TCPAddr).Port)
+	ln.Close()
+	return port, nil
+}
+
+// logWriter adapts log.Entry to io.Writer for piping process output.
+type logWriter struct {
+	entry *log.Entry
+	level log.Level
+}
+
+func (w *logWriter) Write(p []byte) (int, error) {
+	msg := strings.TrimRight(string(p), "\n\r")
+	if msg == "" {
+		return len(p), nil
+	}
+	switch w.level {
+	case log.WarnLevel:
+		w.entry.Warn(msg)
+	default:
+		w.entry.Debug(msg)
+	}
+	return len(p), nil
+}
+
+// Ensure logWriter satisfies io.Writer.
+var _ io.Writer = (*logWriter)(nil)

client/inspect/envoy_config.go

@@ -0,0 +1,382 @@
+package inspect
+
+import (
+	"bytes"
+	"fmt"
+	"strings"
+	"text/template"
+)
+
+// envoyBootstrapTmpl generates the full envoy bootstrap with rule translation.
+// TLS rules become per-SNI filter chains; HTTP rules become per-domain virtual hosts.
+var envoyBootstrapTmpl = template.Must(template.New("bootstrap").Funcs(template.FuncMap{
+	"quote": func(s string) string { return fmt.Sprintf("%q", s) },
+}).Parse(`node:
+  id: netbird-inspect
+  cluster: netbird
+admin:
+  address:
+    socket_address:
+      address: 127.0.0.1
+      port_value: {{.AdminPort}}
+static_resources:
+  listeners:
+    - name: inspect_listener
+      address:
+        socket_address:
+          address: 127.0.0.1
+          port_value: {{.ListenPort}}
+      listener_filters:
+        - name: envoy.filters.listener.proxy_protocol
+          typed_config:
+            "@type": type.googleapis.com/envoy.extensions.filters.listener.proxy_protocol.v3.ProxyProtocol
+        - name: envoy.filters.listener.tls_inspector
+          typed_config:
+            "@type": type.googleapis.com/envoy.extensions.filters.listener.tls_inspector.v3.TlsInspector
+      filter_chains:
+{{- /* TLS filter chains: per-SNI block/allow + default */ -}}
+{{- range .TLSChains}}
+        - filter_chain_match:
+            transport_protocol: tls
+{{- if .ServerNames}}
+            server_names:
+{{- range .ServerNames}}
+              - {{quote .}}
+{{- end}}
+{{- end}}
+          filters:
+{{$.NetworkFiltersSnippet}}            - name: envoy.filters.network.tcp_proxy
+              typed_config:
+                "@type": type.googleapis.com/envoy.extensions.filters.network.tcp_proxy.v3.TcpProxy
+                stat_prefix: {{.StatPrefix}}
+                cluster: original_dst
+                access_log:
+                  - name: envoy.access_loggers.stderr
+                    typed_config:
+                      "@type": type.googleapis.com/envoy.extensions.access_loggers.stream.v3.StderrAccessLog
+                      log_format:
+                        text_format: "[%START_TIME%] tcp %DOWNSTREAM_REMOTE_ADDRESS% -> %UPSTREAM_HOST% %RESPONSE_FLAGS% %DURATION%ms\n"
+{{- end}}
+{{- /* Plain HTTP filter chain with per-domain virtual hosts */}}
+        - filters:
+            - name: envoy.filters.network.http_connection_manager
+              typed_config:
+                "@type": type.googleapis.com/envoy.extensions.filters.network.http_connection_manager.v3.HttpConnectionManager
+                stat_prefix: inspect_http
+                access_log:
+                  - name: envoy.access_loggers.stderr
+                    typed_config:
+                      "@type": type.googleapis.com/envoy.extensions.access_loggers.stream.v3.StderrAccessLog
+                      log_format:
+                        text_format: "[%START_TIME%] http %DOWNSTREAM_REMOTE_ADDRESS% %REQ(:AUTHORITY)% %REQ(:METHOD)% %REQ(X-ENVOY-ORIGINAL-PATH?:PATH)% %RESPONSE_CODE% %RESPONSE_FLAGS% %DURATION%ms\n"
+                http_filters:
+{{.HTTPFiltersSnippet}}                  - name: envoy.filters.http.router
+                    typed_config:
+                      "@type": type.googleapis.com/envoy.extensions.filters.http.router.v3.Router
+                route_config:
+                  virtual_hosts:
+{{- range .VirtualHosts}}
+                    - name: {{.Name}}
+                      domains: [{{.DomainsStr}}]
+                      routes:
+{{- range .Routes}}
+                        - match:
+                            prefix: "{{if .PathPrefix}}{{.PathPrefix}}{{else}}/{{end}}"
+{{- if .Block}}
+                          direct_response:
+                            status: 403
+                            body:
+                              filename: "{{$.BlockPagePath}}"
+{{- else}}
+                          route:
+                            cluster: original_dst
+{{- end}}
+{{- end}}
+{{- end}}
+  clusters:
+    - name: original_dst
+      type: ORIGINAL_DST
+      lb_policy: CLUSTER_PROVIDED
+      connect_timeout: 10s
+{{.ExtraClusters}}`))
+
+// tlsChain represents a TLS filter chain entry for the template.
+// All TLS chains are passthrough (block decisions happen in Go before envoy).
+type tlsChain struct {
+	// ServerNames restricts this chain to specific SNIs. Empty is catch-all.
+	ServerNames []string
+	StatPrefix  string
+}
+
+// envoyRoute represents a single route entry within a virtual host.
+type envoyRoute struct {
+	// PathPrefix for envoy prefix match. Empty means catch-all "/".
+	PathPrefix string
+	Block      bool
+}
+
+// virtualHost represents an HTTP virtual host entry for the template.
+type virtualHost struct {
+	Name string
+	// DomainsStr is pre-formatted for the template: "a", "b".
+	DomainsStr string
+	Routes     []envoyRoute
+}
+
+type bootstrapData struct {
+	AdminPort             uint16
+	ListenPort            uint16
+	BlockPagePath         string
+	TLSChains             []tlsChain
+	VirtualHosts          []virtualHost
+	HTTPFiltersSnippet    string
+	NetworkFiltersSnippet string
+	ExtraClusters         string
+}
+
+// generateBootstrap produces the envoy bootstrap YAML from the inspect config.
+// Translates inspection rules into envoy-native per-SNI and per-domain routing.
+// blockPagePath is the path to the HTML block page file served by direct_response.
+func generateBootstrap(config Config, listenPort, adminPort uint16, blockPagePath string) ([]byte, error) {
+	data := bootstrapData{
+		AdminPort:     adminPort,
+		BlockPagePath: blockPagePath,
+		ListenPort:    listenPort,
+		TLSChains:     buildTLSChains(config),
+		VirtualHosts:  buildVirtualHosts(config),
+	}
+
+	if config.Envoy != nil && config.Envoy.Snippets != nil {
+		s := config.Envoy.Snippets
+		data.HTTPFiltersSnippet = indentSnippet(s.HTTPFilters, 18)
+		data.NetworkFiltersSnippet = indentSnippet(s.NetworkFilters, 12)
+		data.ExtraClusters = indentSnippet(s.Clusters, 4)
+	}
+
+	var buf bytes.Buffer
+	if err := envoyBootstrapTmpl.Execute(&buf, data); err != nil {
+		return nil, fmt.Errorf("execute bootstrap template: %w", err)
+	}
+
+	return buf.Bytes(), nil
+}
+
+// buildTLSChains translates inspection rules into envoy TLS filter chains.
+// Block rules -> per-SNI chain routing to blackhole.
+// Allow rules (when default=block) -> per-SNI chain routing to original_dst.
+// Default chain follows DefaultAction.
+func buildTLSChains(config Config) []tlsChain {
+	// TLS block decisions happen in Go before forwarding to envoy, so we only
+	// generate allow/passthrough chains here. Envoy can't cleanly close a TLS
+	// connection without completing a handshake, so blocked SNIs never reach envoy.
+	var allowed []string
+
+	for _, rule := range config.Rules {
+		if !ruleTouchesProtocol(rule, ProtoHTTPS, ProtoH2) {
+			continue
+		}
+		for _, d := range rule.Domains {
+			sni := d.PunycodeString()
+			if rule.Action == ActionAllow || rule.Action == ActionInspect {
+				allowed = append(allowed, sni)
+			}
+		}
+	}
+
+	var chains []tlsChain
+
+	if len(allowed) > 0 && config.DefaultAction == ActionBlock {
+		chains = append(chains, tlsChain{
+			ServerNames: allowed,
+			StatPrefix:  "tls_allowed",
+		})
+	}
+
+	// Default catch-all: passthrough (blocked SNIs never arrive here)
+	chains = append(chains, tlsChain{
+		StatPrefix: "tls_default",
+	})
+
+	return chains
+}
+
+// buildVirtualHosts translates inspection rules into envoy HTTP virtual hosts.
+// Groups rules by domain, generates per-path routes within each virtual host.
+func buildVirtualHosts(config Config) []virtualHost {
+	// Group rules by domain for per-domain virtual hosts.
+	type domainRules struct {
+		domains []string
+		routes  []envoyRoute
+	}
+
+	domainRouteMap := make(map[string][]envoyRoute)
+
+	for _, rule := range config.Rules {
+		if !ruleTouchesProtocol(rule, ProtoHTTP, ProtoWebSocket) {
+			continue
+		}
+		isBlock := rule.Action == ActionBlock
+
+		// Rules without domains or paths are handled by the default action.
+		if len(rule.Domains) == 0 && len(rule.Paths) == 0 {
+			continue
+		}
+
+		// Build routes for this rule's paths
+		var routes []envoyRoute
+		if len(rule.Paths) > 0 {
+			for _, p := range rule.Paths {
+				// Convert our path patterns to envoy prefix match.
+				// Strip trailing * for envoy prefix matching.
+				prefix := strings.TrimSuffix(p, "*")
+				routes = append(routes, envoyRoute{PathPrefix: prefix, Block: isBlock})
+			}
+		} else {
+			routes = append(routes, envoyRoute{Block: isBlock})
+		}
+
+		if len(rule.Domains) > 0 {
+			for _, d := range rule.Domains {
+				host := d.PunycodeString()
+				domainRouteMap[host] = append(domainRouteMap[host], routes...)
+			}
+		} else {
+			// No domain: applies to all, add to default host
+			domainRouteMap["*"] = append(domainRouteMap["*"], routes...)
+		}
+	}
+
+	var hosts []virtualHost
+	idx := 0
+
+	// Per-domain virtual hosts with path routes
+	for domain, routes := range domainRouteMap {
+		if domain == "*" {
+			continue
+		}
+		// Add a catch-all route after path-specific routes.
+		// The catch-all follows the default action.
+		routes = append(routes, envoyRoute{Block: config.DefaultAction == ActionBlock})
+
+		hosts = append(hosts, virtualHost{
+			Name:       fmt.Sprintf("domain_%d", idx),
+			DomainsStr: fmt.Sprintf("%q", domain),
+			Routes:     routes,
+		})
+		idx++
+	}
+
+	// Default virtual host (catch-all for unmatched domains)
+	defaultRoutes := domainRouteMap["*"]
+	defaultRoutes = append(defaultRoutes, envoyRoute{Block: config.DefaultAction == ActionBlock})
+	hosts = append(hosts, virtualHost{
+		Name:       "default",
+		DomainsStr: `"*"`,
+		Routes:     defaultRoutes,
+	})
+
+	return hosts
+}
+
+// ruleTouchesProtocol returns true if the rule's protocol list includes any of the given protocols,
+// or if the protocol list is empty (matches all).
+func ruleTouchesProtocol(rule Rule, protos ...ProtoType) bool {
+	if len(rule.Protocols) == 0 {
+		return true
+	}
+	for _, rp := range rule.Protocols {
+		for _, p := range protos {
+			if rp == p {
+				return true
+			}
+		}
+	}
+	return false
+}
+
+// indentSnippet prepends each line of the YAML snippet with the given number of spaces.
+// Returns empty string if snippet is empty.
+func indentSnippet(snippet string, spaces int) string {
+	if snippet == "" {
+		return ""
+	}
+
+	prefix := make([]byte, spaces)
+	for i := range prefix {
+		prefix[i] = ' '
+	}
+
+	var buf bytes.Buffer
+	for i, line := range bytes.Split([]byte(snippet), []byte("\n")) {
+		if i > 0 {
+			buf.WriteByte('\n')
+		}
+		if len(line) > 0 {
+			buf.Write(prefix)
+			buf.Write(line)
+		}
+	}
+	buf.WriteByte('\n')
+
+	return buf.String()
+}
+
+// ValidateSnippets checks that user-provided snippets are safe to inject
+// into the envoy config. Returns an error describing the first violation found.
+//
+// Validation rules:
+//   - Each snippet must be valid YAML (prevents syntax-level injection)
+//   - Snippets must not contain YAML document separators (--- or ...) that could
+//     break out of the indentation context
+//   - Snippets must only contain list items (starting with "- ") at the top level,
+//     matching what envoy expects for filters and clusters
+func ValidateSnippets(snippets *EnvoySnippets) error {
+	if snippets == nil {
+		return nil
+	}
+
+	fields := []struct {
+		name  string
+		value string
+	}{
+		{"http_filters", snippets.HTTPFilters},
+		{"network_filters", snippets.NetworkFilters},
+		{"clusters", snippets.Clusters},
+	}
+
+	for _, f := range fields {
+		if f.value == "" {
+			continue
+		}
+		if err := validateSnippetYAML(f.name, f.value); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+func validateSnippetYAML(name, snippet string) error {
+	// Check for YAML document markers that could break template structure.
+	for _, line := range strings.Split(snippet, "\n") {
+		trimmed := strings.TrimSpace(line)
+		if trimmed == "---" || trimmed == "..." {
+			return fmt.Errorf("snippet %q: YAML document separators (--- or ...) are not allowed", name)
+		}
+	}
+
+	// Verify it's valid YAML by checking it doesn't cause template execution issues.
+	// We can't import yaml.v3 here without adding a dependency, so we do structural checks.
+
+	// Check for null bytes or control characters that could confuse YAML parsers.
+	for i, b := range []byte(snippet) {
+		if b == 0 {
+			return fmt.Errorf("snippet %q: null byte at position %d", name, i)
+		}
+		if b < 0x09 || (b > 0x0D && b < 0x20 && b != 0x1B) {
+			return fmt.Errorf("snippet %q: control character 0x%02x at position %d", name, b, i)
+		}
+	}
+
+	return nil
+}

client/inspect/envoy_forward.go

@@ -0,0 +1,88 @@
+package inspect
+
+import (
+	"context"
+	"encoding/binary"
+	"fmt"
+	"net"
+	"net/netip"
+)
+
+// PROXY protocol v2 constants (RFC 7239 / HAProxy spec)
+var proxyV2Signature = [12]byte{
+	0x0D, 0x0A, 0x0D, 0x0A, 0x00, 0x0D, 0x0A, 0x51,
+	0x55, 0x49, 0x54, 0x0A,
+}
+
+const (
+	proxyV2VersionCommand = 0x21 // version 2, PROXY command
+	proxyV2FamilyTCP4     = 0x11 // AF_INET, STREAM
+	proxyV2FamilyTCP6     = 0x21 // AF_INET6, STREAM
+)
+
+// forwardToEnvoy forwards a connection to the given envoy sidecar via PROXY protocol v2.
+// The caller provides the envoy manager snapshot to avoid accessing p.envoy without lock.
+func (p *Proxy) forwardToEnvoy(ctx context.Context, pconn *peekConn, dst netip.AddrPort, src SourceInfo, em *envoyManager) error {
+	envoyAddr := em.ListenAddr()
+
+	conn, err := (&net.Dialer{}).DialContext(ctx, "tcp", envoyAddr.String())
+	if err != nil {
+		return fmt.Errorf("dial envoy at %s: %w", envoyAddr, err)
+	}
+	defer func() {
+		if err := conn.Close(); err != nil {
+			p.log.Debugf("close envoy conn: %v", err)
+		}
+	}()
+
+	if err := writeProxyV2Header(conn, src.IP, dst); err != nil {
+		return fmt.Errorf("write PROXY v2 header: %w", err)
+	}
+
+	p.log.Tracef("envoy: forwarded %s -> %s via PROXY v2", src.IP, dst)
+
+	return relay(ctx, pconn, conn)
+}
+
+// writeProxyV2Header writes a PROXY protocol v2 header to w.
+// The header encodes the original source IP and the destination address:port.
+func writeProxyV2Header(w net.Conn, srcIP netip.Addr, dst netip.AddrPort) error {
+	srcIP = srcIP.Unmap()
+	dstIP := dst.Addr().Unmap()
+
+	var (
+		family byte
+		addrs  []byte
+	)
+
+	if srcIP.Is4() && dstIP.Is4() {
+		family = proxyV2FamilyTCP4
+		s4 := srcIP.As4()
+		d4 := dstIP.As4()
+		addrs = make([]byte, 12) // 4+4+2+2
+		copy(addrs[0:4], s4[:])
+		copy(addrs[4:8], d4[:])
+		binary.BigEndian.PutUint16(addrs[8:10], 0) // src port unknown
+		binary.BigEndian.PutUint16(addrs[10:12], dst.Port())
+	} else {
+		family = proxyV2FamilyTCP6
+		s16 := srcIP.As16()
+		d16 := dstIP.As16()
+		addrs = make([]byte, 36) // 16+16+2+2
+		copy(addrs[0:16], s16[:])
+		copy(addrs[16:32], d16[:])
+		binary.BigEndian.PutUint16(addrs[32:34], 0) // src port unknown
+		binary.BigEndian.PutUint16(addrs[34:36], dst.Port())
+	}
+
+	// Header: signature(12) + ver_cmd(1) + family(1) + len(2) + addrs
+	header := make([]byte, 16+len(addrs))
+	copy(header[0:12], proxyV2Signature[:])
+	header[12] = proxyV2VersionCommand
+	header[13] = family
+	binary.BigEndian.PutUint16(header[14:16], uint16(len(addrs)))
+	copy(header[16:], addrs)
+
+	_, err := w.Write(header)
+	return err
+}

client/inspect/envoy_signal.go

@@ -0,0 +1,13 @@
+//go:build !windows
+
+package inspect
+
+import (
+	"os"
+	"syscall"
+)
+
+// signalReload sends SIGHUP to the envoy process to trigger config reload.
+func signalReload(p *os.Process) error {
+	return p.Signal(syscall.SIGHUP)
+}

client/inspect/envoy_signal_windows.go

@@ -0,0 +1,13 @@
+//go:build windows
+
+package inspect
+
+import (
+	"fmt"
+	"os"
+)
+
+// signalReload is not supported on Windows. Envoy must be restarted.
+func signalReload(_ *os.Process) error {
+	return fmt.Errorf("envoy config reload via signal not supported on Windows")
+}

client/inspect/external.go

@@ -0,0 +1,229 @@
+package inspect
+
+import (
+	"bufio"
+	"context"
+	"encoding/base64"
+	"fmt"
+	"io"
+	"net"
+	"net/http"
+	"net/netip"
+	"net/url"
+	"time"
+)
+
+const (
+	externalDialTimeout = 10 * time.Second
+)
+
+// handleExternal forwards the connection to an external proxy.
+// For TLS connections, it uses HTTP CONNECT to tunnel through the proxy.
+// For HTTP connections, it rewrites the request to use the proxy.
+func (p *Proxy) handleExternal(ctx context.Context, pconn *peekConn, dst netip.AddrPort) error {
+	p.mu.RLock()
+	proxyURL := p.config.ExternalURL
+	p.mu.RUnlock()
+
+	if proxyURL == nil {
+		return fmt.Errorf("external proxy URL not configured")
+	}
+
+	switch proxyURL.Scheme {
+	case "http", "https":
+		return p.externalHTTPProxy(ctx, pconn, dst, proxyURL)
+	case "socks5":
+		return p.externalSOCKS5(ctx, pconn, dst, proxyURL)
+	default:
+		return fmt.Errorf("unsupported external proxy scheme: %s", proxyURL.Scheme)
+	}
+}
+
+// externalHTTPProxy tunnels through an HTTP proxy using CONNECT.
+func (p *Proxy) externalHTTPProxy(ctx context.Context, pconn *peekConn, dst netip.AddrPort, proxyURL *url.URL) error {
+	proxyAddr := proxyURL.Host
+	if _, _, err := net.SplitHostPort(proxyAddr); err != nil {
+		proxyAddr = net.JoinHostPort(proxyAddr, "8080")
+	}
+
+	proxyConn, err := (&net.Dialer{Timeout: externalDialTimeout}).DialContext(ctx, "tcp", proxyAddr)
+	if err != nil {
+		return fmt.Errorf("dial external proxy %s: %w", proxyAddr, err)
+	}
+	defer func() {
+		if err := proxyConn.Close(); err != nil {
+			p.log.Debugf("close external proxy conn: %v", err)
+		}
+	}()
+
+	connectReq := fmt.Sprintf("CONNECT %s HTTP/1.1\r\nHost: %s\r\n", dst.String(), dst.String())
+	if proxyURL.User != nil {
+		connectReq += "Proxy-Authorization: Basic " + basicAuth(proxyURL.User) + "\r\n"
+	}
+	connectReq += "\r\n"
+
+	if _, err := io.WriteString(proxyConn, connectReq); err != nil {
+		return fmt.Errorf("send CONNECT to proxy: %w", err)
+	}
+
+	resp, err := http.ReadResponse(bufio.NewReader(proxyConn), nil)
+	if err != nil {
+		return fmt.Errorf("read CONNECT response: %w", err)
+	}
+	if err := resp.Body.Close(); err != nil {
+		p.log.Debugf("close CONNECT resp body: %v", err)
+	}
+
+	if resp.StatusCode != http.StatusOK {
+		return fmt.Errorf("proxy CONNECT failed: %s", resp.Status)
+	}
+
+	return relay(ctx, pconn, proxyConn)
+}
+
+// externalSOCKS5 tunnels through a SOCKS5 proxy.
+func (p *Proxy) externalSOCKS5(ctx context.Context, pconn *peekConn, dst netip.AddrPort, proxyURL *url.URL) error {
+	proxyAddr := proxyURL.Host
+	if _, _, err := net.SplitHostPort(proxyAddr); err != nil {
+		proxyAddr = net.JoinHostPort(proxyAddr, "1080")
+	}
+
+	proxyConn, err := (&net.Dialer{Timeout: externalDialTimeout}).DialContext(ctx, "tcp", proxyAddr)
+	if err != nil {
+		return fmt.Errorf("dial SOCKS5 proxy %s: %w", proxyAddr, err)
+	}
+	defer func() {
+		if err := proxyConn.Close(); err != nil {
+			p.log.Debugf("close SOCKS5 proxy conn: %v", err)
+		}
+	}()
+
+	if err := socks5Handshake(proxyConn, dst, proxyURL.User); err != nil {
+		return fmt.Errorf("SOCKS5 handshake: %w", err)
+	}
+
+	return relay(ctx, pconn, proxyConn)
+}
+
+// socks5Handshake performs the SOCKS5 handshake to connect through the proxy.
+func socks5Handshake(conn net.Conn, dst netip.AddrPort, userinfo *url.Userinfo) error {
+	needAuth := userinfo != nil
+
+	// Greeting
+	var methods []byte
+	if needAuth {
+		methods = []byte{0x00, 0x02} // no auth, username/password
+	} else {
+		methods = []byte{0x00} // no auth
+	}
+	greeting := append([]byte{0x05, byte(len(methods))}, methods...)
+	if _, err := conn.Write(greeting); err != nil {
+		return fmt.Errorf("send greeting: %w", err)
+	}
+
+	// Server method selection
+	var methodResp [2]byte
+	if _, err := io.ReadFull(conn, methodResp[:]); err != nil {
+		return fmt.Errorf("read method selection: %w", err)
+	}
+	if methodResp[0] != 0x05 {
+		return fmt.Errorf("unexpected SOCKS version: %d", methodResp[0])
+	}
+
+	// Handle authentication if selected
+	if methodResp[1] == 0x02 {
+		if err := socks5Auth(conn, userinfo); err != nil {
+			return err
+		}
+	} else if methodResp[1] != 0x00 {
+		return fmt.Errorf("unsupported SOCKS5 auth method: %d", methodResp[1])
+	}
+
+	// Connection request
+	addr := dst.Addr()
+	var addrBytes []byte
+	if addr.Is4() {
+		a4 := addr.As4()
+		addrBytes = append([]byte{0x01}, a4[:]...) // IPv4
+	} else {
+		a16 := addr.As16()
+		addrBytes = append([]byte{0x04}, a16[:]...) // IPv6
+	}
+
+	port := dst.Port()
+	connectReq := append([]byte{0x05, 0x01, 0x00}, addrBytes...)
+	connectReq = append(connectReq, byte(port>>8), byte(port))
+
+	if _, err := conn.Write(connectReq); err != nil {
+		return fmt.Errorf("send connect request: %w", err)
+	}
+
+	// Read response (minimum 10 bytes for IPv4)
+	var respHeader [4]byte
+	if _, err := io.ReadFull(conn, respHeader[:]); err != nil {
+		return fmt.Errorf("read connect response: %w", err)
+	}
+	if respHeader[1] != 0x00 {
+		return fmt.Errorf("SOCKS5 connect failed: status %d", respHeader[1])
+	}
+
+	// Skip bound address
+	switch respHeader[3] {
+	case 0x01: // IPv4
+		var skip [4 + 2]byte
+		if _, err := io.ReadFull(conn, skip[:]); err != nil {
+			return fmt.Errorf("read SOCKS5 bound IPv4 address: %w", err)
+		}
+	case 0x04: // IPv6
+		var skip [16 + 2]byte
+		if _, err := io.ReadFull(conn, skip[:]); err != nil {
+			return fmt.Errorf("read SOCKS5 bound IPv6 address: %w", err)
+		}
+	case 0x03: // Domain
+		var dLen [1]byte
+		if _, err := io.ReadFull(conn, dLen[:]); err != nil {
+			return fmt.Errorf("read domain length: %w", err)
+		}
+		skip := make([]byte, int(dLen[0])+2)
+		if _, err := io.ReadFull(conn, skip); err != nil {
+			return fmt.Errorf("read SOCKS5 bound domain address: %w", err)
+		}
+	}
+
+	return nil
+}
+
+func socks5Auth(conn net.Conn, userinfo *url.Userinfo) error {
+	if userinfo == nil {
+		return fmt.Errorf("SOCKS5 auth required but no credentials provided")
+	}
+
+	user := userinfo.Username()
+	pass, _ := userinfo.Password()
+
+	// Username/password auth (RFC 1929)
+	auth := []byte{0x01, byte(len(user))}
+	auth = append(auth, []byte(user)...)
+	auth = append(auth, byte(len(pass)))
+	auth = append(auth, []byte(pass)...)
+
+	if _, err := conn.Write(auth); err != nil {
+		return fmt.Errorf("send auth: %w", err)
+	}
+
+	var resp [2]byte
+	if _, err := io.ReadFull(conn, resp[:]); err != nil {
+		return fmt.Errorf("read auth response: %w", err)
+	}
+	if resp[1] != 0x00 {
+		return fmt.Errorf("SOCKS5 auth failed: status %d", resp[1])
+	}
+
+	return nil
+}
+
+func basicAuth(userinfo *url.Userinfo) string {
+	user := userinfo.Username()
+	pass, _ := userinfo.Password()
+	return base64.StdEncoding.EncodeToString([]byte(user + ":" + pass))
+}

client/inspect/http.go

@@ -0,0 +1,532 @@
+package inspect
+
+import (
+	"bufio"
+	"context"
+	"fmt"
+	"io"
+	"net"
+	"net/http"
+	"net/netip"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/netbirdio/netbird/shared/management/domain"
+)
+
+const (
+	headerUpgrade  = "Upgrade"
+	valueWebSocket = "websocket"
+)
+
+// inspectHTTP runs the HTTP inspection pipeline on decrypted traffic.
+// It handles HTTP/1.1 (request-response loop), HTTP/2 (via Go stdlib reverse proxy),
+// and WebSocket upgrade detection.
+func (p *Proxy) inspectHTTP(ctx context.Context, client, remote net.Conn, dst netip.AddrPort, sni domain.Domain, src SourceInfo, proto string) error {
+	if proto == "h2" {
+		return p.inspectH2(ctx, client, remote, dst, sni, src)
+	}
+	return p.inspectH1(ctx, client, remote, dst, sni, src)
+}
+
+// inspectH1 handles HTTP/1.1 request-response inspection in a loop.
+func (p *Proxy) inspectH1(ctx context.Context, client, remote net.Conn, dst netip.AddrPort, sni domain.Domain, src SourceInfo) error {
+	clientReader := bufio.NewReader(client)
+	remoteReader := bufio.NewReader(remote)
+
+	for {
+		if ctx.Err() != nil {
+			return ctx.Err()
+		}
+
+		// Set idle timeout between requests to prevent connection hogging.
+		if err := client.SetReadDeadline(time.Now().Add(idleTimeout)); err != nil {
+			return fmt.Errorf("set idle deadline: %w", err)
+		}
+		req, err := http.ReadRequest(clientReader)
+		if err != nil {
+			if isClosedErr(err) {
+				return nil
+			}
+			return fmt.Errorf("read HTTP request: %w", err)
+		}
+		if err := client.SetReadDeadline(time.Time{}); err != nil {
+			return fmt.Errorf("clear read deadline: %w", err)
+		}
+
+		// Re-evaluate rules based on Host header if SNI was empty
+		host := hostFromRequest(req, sni)
+
+		// Domain fronting: Host header doesn't match TLS SNI
+		if isDomainFronting(req, sni) {
+			p.log.Debugf("domain fronting detected: SNI=%s Host=%s", sni.PunycodeString(), host.PunycodeString())
+			writeBlockResponse(client, req, host)
+			return ErrBlocked
+		}
+
+		proto := ProtoHTTP
+		if isWebSocketUpgrade(req) {
+			proto = ProtoWebSocket
+		}
+		action := p.evaluateAction(src.IP, host, dst, proto, req.URL.Path)
+		if action == ActionBlock {
+			p.log.Debugf("block: HTTP %s %s (host=%s)", req.Method, req.URL.Path, host.PunycodeString())
+			writeBlockResponse(client, req, host)
+			return ErrBlocked
+		}
+		p.log.Tracef("allow: HTTP %s %s (host=%s, action=%s)", req.Method, req.URL.Path, host.PunycodeString(), action)
+
+		// ICAP REQMOD: send request for inspection.
+		// Snapshot ICAP client under lock to avoid use-after-close races.
+		p.mu.RLock()
+		icap := p.icap
+		p.mu.RUnlock()
+		if icap != nil {
+			modified, err := icap.ReqMod(req)
+			if err != nil {
+				p.log.Debugf("ICAP REQMOD error for %s: %v", host.PunycodeString(), err)
+				// Fail-closed: block on ICAP error
+				writeBlockResponse(client, req, host)
+				return fmt.Errorf("ICAP REQMOD: %w", err)
+			}
+			req = modified
+		}
+
+		if isWebSocketUpgrade(req) {
+			return p.handleWebSocket(ctx, req, client, clientReader, remote, remoteReader)
+		}
+
+		removeHopByHopHeaders(req.Header)
+
+		if err := req.Write(remote); err != nil {
+			return fmt.Errorf("forward request: %w", err)
+		}
+
+		resp, err := http.ReadResponse(remoteReader, req)
+		if err != nil {
+			return fmt.Errorf("read HTTP response: %w", err)
+		}
+
+		// ICAP RESPMOD: send response for inspection
+		if icap != nil {
+			modified, err := icap.RespMod(req, resp)
+			if err != nil {
+				p.log.Debugf("ICAP RESPMOD error for %s: %v", host.PunycodeString(), err)
+				if err := resp.Body.Close(); err != nil {
+					p.log.Debugf("close resp body: %v", err)
+				}
+				writeBlockResponse(client, req, host)
+				return fmt.Errorf("ICAP RESPMOD: %w", err)
+			}
+			resp = modified
+		}
+
+		removeHopByHopHeaders(resp.Header)
+
+		if err := resp.Write(client); err != nil {
+			if closeErr := resp.Body.Close(); closeErr != nil {
+				p.log.Debugf("close resp body: %v", closeErr)
+			}
+			return fmt.Errorf("forward response: %w", err)
+		}
+		if err := resp.Body.Close(); err != nil {
+			p.log.Debugf("close resp body: %v", err)
+		}
+
+		// Connection: close means we're done
+		if resp.Close || req.Close {
+			return nil
+		}
+	}
+}
+
+// inspectH2 proxies HTTP/2 traffic using Go's http stack.
+// Client and remote are already-established TLS connections with h2 negotiated.
+func (p *Proxy) inspectH2(ctx context.Context, client, remote net.Conn, dst netip.AddrPort, sni domain.Domain, src SourceInfo) error {
+	// For h2 MITM inspection, we use a local http.Server reading from the client
+	// connection and an http.Transport writing to the remote connection.
+	//
+	// The transport is configured to use the existing TLS connection to the
+	// real server. The handler inspects each request/response pair.
+
+	transport := &http.Transport{
+		DialContext: func(_ context.Context, _, _ string) (net.Conn, error) {
+			return remote, nil
+		},
+		DialTLSContext: func(_ context.Context, _, _ string) (net.Conn, error) {
+			return remote, nil
+		},
+		ForceAttemptHTTP2: true,
+	}
+
+	handler := &h2InspectionHandler{
+		proxy:     p,
+		transport: transport,
+		dst:       dst,
+		sni:       sni,
+		src:       src,
+	}
+
+	server := &http.Server{
+		Handler: handler,
+	}
+
+	// Serve the single client connection.
+	// ServeConn blocks until the connection is done.
+	errCh := make(chan error, 1)
+	go func() {
+		// http.Server doesn't have a direct ServeConn for h2,
+		// so we use Serve with a single-connection listener.
+		ln := &singleConnListener{conn: client}
+		errCh <- server.Serve(ln)
+	}()
+
+	select {
+	case <-ctx.Done():
+		if err := server.Close(); err != nil {
+			p.log.Debugf("close h2 server: %v", err)
+		}
+		return ctx.Err()
+	case err := <-errCh:
+		if err == http.ErrServerClosed {
+			return nil
+		}
+		return err
+	}
+}
+
+// h2InspectionHandler inspects each HTTP/2 request/response pair.
+type h2InspectionHandler struct {
+	proxy     *Proxy
+	transport http.RoundTripper
+	dst       netip.AddrPort
+	sni       domain.Domain
+	src       SourceInfo
+}
+
+func (h *h2InspectionHandler) ServeHTTP(w http.ResponseWriter, req *http.Request) {
+	host := hostFromRequest(req, h.sni)
+
+	if isDomainFronting(req, h.sni) {
+		h.proxy.log.Debugf("domain fronting detected: SNI=%s Host=%s", h.sni.PunycodeString(), host.PunycodeString())
+		writeBlockPage(w, host)
+		return
+	}
+
+	action := h.proxy.evaluateAction(h.src.IP, host, h.dst, ProtoH2, req.URL.Path)
+	if action == ActionBlock {
+		h.proxy.log.Debugf("block: H2 %s %s (host=%s)", req.Method, req.URL.Path, host.PunycodeString())
+		writeBlockPage(w, host)
+		return
+	}
+
+	// ICAP REQMOD
+	if h.proxy.icap != nil {
+		modified, err := h.proxy.icap.ReqMod(req)
+		if err != nil {
+			h.proxy.log.Debugf("ICAP REQMOD error for %s: %v", host.PunycodeString(), err)
+			writeBlockPage(w, host)
+			return
+		}
+		req = modified
+	}
+
+	// Forward to upstream
+	req.URL.Scheme = "https"
+	req.URL.Host = h.sni.PunycodeString()
+	req.RequestURI = ""
+
+	resp, err := h.transport.RoundTrip(req)
+	if err != nil {
+		h.proxy.log.Debugf("h2 upstream error for %s: %v", host.PunycodeString(), err)
+		http.Error(w, "Bad Gateway", http.StatusBadGateway)
+		return
+	}
+	defer func() {
+		if err := resp.Body.Close(); err != nil {
+			h.proxy.log.Debugf("close h2 resp body: %v", err)
+		}
+	}()
+
+	// ICAP RESPMOD
+	if h.proxy.icap != nil {
+		modified, err := h.proxy.icap.RespMod(req, resp)
+		if err != nil {
+			h.proxy.log.Debugf("ICAP RESPMOD error for %s: %v", host.PunycodeString(), err)
+			writeBlockPage(w, host)
+			return
+		}
+		resp = modified
+	}
+
+	// Copy response headers and body
+	for k, vals := range resp.Header {
+		for _, v := range vals {
+			w.Header().Add(k, v)
+		}
+	}
+	w.WriteHeader(resp.StatusCode)
+	if _, err := io.Copy(w, resp.Body); err != nil {
+		h.proxy.log.Debugf("h2 response copy error: %v", err)
+	}
+}
+
+// handleWebSocket completes the WebSocket upgrade and relays frames bidirectionally.
+func (p *Proxy) handleWebSocket(ctx context.Context, req *http.Request, client io.ReadWriter, clientReader *bufio.Reader, remote io.ReadWriter, remoteReader *bufio.Reader) error {
+	if err := req.Write(remote); err != nil {
+		return fmt.Errorf("forward WebSocket upgrade: %w", err)
+	}
+
+	resp, err := http.ReadResponse(remoteReader, req)
+	if err != nil {
+		return fmt.Errorf("read WebSocket upgrade response: %w", err)
+	}
+
+	if err := resp.Write(client); err != nil {
+		if closeErr := resp.Body.Close(); closeErr != nil {
+			p.log.Debugf("close ws resp body: %v", closeErr)
+		}
+		return fmt.Errorf("forward WebSocket upgrade response: %w", err)
+	}
+	if err := resp.Body.Close(); err != nil {
+		p.log.Debugf("close ws resp body: %v", err)
+	}
+
+	if resp.StatusCode != http.StatusSwitchingProtocols {
+		return fmt.Errorf("WebSocket upgrade rejected: status %d", resp.StatusCode)
+	}
+
+	p.log.Tracef("allow: WebSocket upgrade for %s", req.Host)
+
+	// Relay WebSocket frames bidirectionally.
+	// clientReader/remoteReader may have buffered data.
+	clientConn := mergeReadWriter(clientReader, client)
+	remoteConn := mergeReadWriter(remoteReader, remote)
+
+	return relayRW(ctx, clientConn, remoteConn)
+}
+
+// hostFromRequest extracts a domain.Domain from the HTTP request Host header,
+// falling back to the SNI if Host is empty or an IP.
+func hostFromRequest(req *http.Request, fallback domain.Domain) domain.Domain {
+	host := req.Host
+	if host == "" {
+		return fallback
+	}
+
+	// Strip port if present
+	if h, _, err := net.SplitHostPort(host); err == nil {
+		host = h
+	}
+
+	// If it's an IP address, use the SNI fallback
+	if _, err := netip.ParseAddr(host); err == nil {
+		return fallback
+	}
+
+	d, err := domain.FromString(host)
+	if err != nil {
+		return fallback
+	}
+	return d
+}
+
+// isDomainFronting detects domain fronting: the Host header doesn't match the
+// SNI used during the TLS handshake. Only meaningful when SNI is non-empty
+// (i.e., we're in MITM mode and know the original SNI).
+func isDomainFronting(req *http.Request, sni domain.Domain) bool {
+	if sni == "" {
+		return false
+	}
+
+	host := hostFromRequest(req, "")
+	if host == "" {
+		return false
+	}
+
+	// Host should match SNI or be a subdomain of SNI
+	if host == sni {
+		return false
+	}
+
+	// Allow www.example.com when SNI is example.com
+	sniStr := sni.PunycodeString()
+	hostStr := host.PunycodeString()
+	if strings.HasSuffix(hostStr, "."+sniStr) {
+		return false
+	}
+
+	return true
+}
+
+func isWebSocketUpgrade(req *http.Request) bool {
+	return strings.EqualFold(req.Header.Get(headerUpgrade), valueWebSocket)
+}
+
+// writeBlockPage writes the styled HTML block page to an http.ResponseWriter (H2 path).
+func writeBlockPage(w http.ResponseWriter, host domain.Domain) {
+	hostname := host.PunycodeString()
+	body := fmt.Sprintf(blockPageHTML, hostname, hostname)
+	w.Header().Set("Content-Type", "text/html; charset=utf-8")
+	w.Header().Set("Cache-Control", "no-store")
+	w.WriteHeader(http.StatusForbidden)
+	io.WriteString(w, body)
+}
+
+func writeBlockResponse(w io.Writer, _ *http.Request, host domain.Domain) {
+	hostname := host.PunycodeString()
+	body := fmt.Sprintf(blockPageHTML, hostname, hostname)
+
+	resp := &http.Response{
+		StatusCode:    http.StatusForbidden,
+		ProtoMajor:    1,
+		ProtoMinor:    1,
+		Header:        make(http.Header),
+		ContentLength: int64(len(body)),
+		Body:          io.NopCloser(strings.NewReader(body)),
+	}
+	resp.Header.Set("Content-Type", "text/html; charset=utf-8")
+	resp.Header.Set("Connection", "close")
+	resp.Header.Set("Cache-Control", "no-store")
+	_ = resp.Write(w)
+}
+
+// blockPageHTML is the self-contained HTML block page.
+// Uses NetBird dark theme with orange accent. Two format args: page title domain, displayed domain.
+const blockPageHTML = `<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Blocked - %s</title>
+<style>
+*{margin:0;padding:0;box-sizing:border-box}
+body{background:#181a1d;color:#d1d5db;font-family:-apple-system,BlinkMacSystemFont,"Segoe UI",Roboto,sans-serif;min-height:100vh;display:flex;align-items:center;justify-content:center}
+.c{text-align:center;max-width:460px;padding:2rem}
+.shield{width:56px;height:56px;margin:0 auto 1.5rem;border-radius:16px;background:#2b2f33;display:flex;align-items:center;justify-content:center}
+.shield svg{width:28px;height:28px;color:#f68330}
+.code{font-size:.8rem;font-weight:500;color:#f68330;font-family:ui-monospace,monospace;letter-spacing:.05em;margin-bottom:.5rem}
+h1{font-size:1.5rem;font-weight:600;color:#f4f4f5;margin-bottom:.5rem}
+p{font-size:.95rem;line-height:1.5;color:#9ca3af;margin-bottom:1.75rem}
+.domain{display:inline-block;background:#25282d;border:1px solid #32363d;border-radius:6px;padding:.15rem .5rem;font-family:ui-monospace,monospace;font-size:.85rem;color:#d1d5db}
+.footer{font-size:.7rem;color:#6b7280;margin-top:2rem;letter-spacing:.03em}
+.footer a{color:#6b7280;text-decoration:none}
+.footer a:hover{color:#9ca3af}
+</style>
+</head>
+<body>
+<div class="c">
+<div class="shield"><svg xmlns="http://www.w3.org/2000/svg" fill="none" viewBox="0 0 24 24" stroke-width="1.5" stroke="currentColor"><path stroke-linecap="round" stroke-linejoin="round" d="M12 9v3.75m0-10.036A11.959 11.959 0 0 1 3.598 6 11.99 11.99 0 0 0 3 9.75c0 5.592 3.824 10.29 9 11.622 5.176-1.332 9-6.03 9-11.622 0-1.31-.21-2.571-.598-3.751A11.96 11.96 0 0 0 12 3.714Z"/></svg></div>
+<div class="code">403 BLOCKED</div>
+<h1>Access Denied</h1>
+<p>This connection to <span class="domain">%s</span> has been blocked by your organization's network policy.</p>
+<div class="footer">Protected by <a href="https://netbird.io" target="_blank" rel="noopener">NetBird</a></div>
+</div>
+</body>
+</html>`
+
+// singleConnListener is a net.Listener that yields a single connection.
+type singleConnListener struct {
+	conn net.Conn
+	once sync.Once
+	ch   chan struct{}
+}
+
+func (l *singleConnListener) Accept() (net.Conn, error) {
+	var accepted bool
+	l.once.Do(func() {
+		l.ch = make(chan struct{})
+		accepted = true
+	})
+	if accepted {
+		return l.conn, nil
+	}
+	// Block until Close
+	<-l.ch
+	return nil, net.ErrClosed
+}
+
+func (l *singleConnListener) Close() error {
+	l.once.Do(func() {
+		l.ch = make(chan struct{})
+	})
+	select {
+	case <-l.ch:
+	default:
+		close(l.ch)
+	}
+	return nil
+}
+
+func (l *singleConnListener) Addr() net.Addr {
+	return l.conn.LocalAddr()
+}
+
+type readWriter struct {
+	io.Reader
+	io.Writer
+}
+
+func mergeReadWriter(r io.Reader, w io.Writer) io.ReadWriter {
+	return &readWriter{Reader: r, Writer: w}
+}
+
+// relayRW copies data bidirectionally between two ReadWriters.
+func relayRW(ctx context.Context, a, b io.ReadWriter) error {
+	ctx, cancel := context.WithCancel(ctx)
+	defer cancel()
+
+	errCh := make(chan error, 2)
+
+	go func() {
+		_, err := io.Copy(b, a)
+		cancel()
+		errCh <- err
+	}()
+
+	go func() {
+		_, err := io.Copy(a, b)
+		cancel()
+		errCh <- err
+	}()
+
+	var firstErr error
+	for range 2 {
+		if err := <-errCh; err != nil && firstErr == nil {
+			if !isClosedErr(err) {
+				firstErr = err
+			}
+		}
+	}
+
+	return firstErr
+}
+
+// hopByHopHeaders are HTTP/1.1 headers that apply to a single connection
+// and must not be forwarded by a proxy (RFC 7230, Section 6.1).
+var hopByHopHeaders = []string{
+	"Connection",
+	"Keep-Alive",
+	"Proxy-Authenticate",
+	"Proxy-Authorization",
+	"TE",
+	"Trailers",
+	"Transfer-Encoding",
+	"Upgrade",
+}
+
+// removeHopByHopHeaders strips hop-by-hop headers from h.
+// Also removes headers listed in the Connection header value.
+func removeHopByHopHeaders(h http.Header) {
+	// First, remove any headers named in the Connection header
+	for _, connHeader := range h["Connection"] {
+		for _, name := range strings.Split(connHeader, ",") {
+			h.Del(strings.TrimSpace(name))
+		}
+	}
+
+	for _, name := range hopByHopHeaders {
+		h.Del(name)
+	}
+}

client/inspect/icap.go

@@ -0,0 +1,479 @@
+package inspect
+
+import (
+	"bufio"
+	"bytes"
+	"fmt"
+	"io"
+	"net"
+	"net/http"
+	"net/textproto"
+	"net/url"
+	"strconv"
+	"strings"
+	"sync"
+	"time"
+
+	log "github.com/sirupsen/logrus"
+)
+
+const (
+	icapVersion     = "ICAP/1.0"
+	icapDefaultPort = "1344"
+	icapConnTimeout = 30 * time.Second
+	icapRWTimeout   = 60 * time.Second
+	icapMaxPoolSize = 8
+	icapIdleTimeout = 60 * time.Second
+	icapMaxRespSize = 4 * 1024 * 1024 // 4 MB
+)
+
+// ICAPClient implements an ICAP (RFC 3507) client with persistent connection pooling.
+type ICAPClient struct {
+	reqModURL  *url.URL
+	respModURL *url.URL
+	pool       chan *icapConn
+	mu         sync.Mutex
+	log        *log.Entry
+	maxPool    int
+}
+
+type icapConn struct {
+	conn    net.Conn
+	reader  *bufio.Reader
+	lastUse time.Time
+}
+
+// NewICAPClient creates an ICAP client. Either or both URLs may be nil
+// to disable that mode.
+func NewICAPClient(logger *log.Entry, cfg *ICAPConfig) *ICAPClient {
+	maxPool := cfg.MaxConnections
+	if maxPool <= 0 {
+		maxPool = icapMaxPoolSize
+	}
+
+	return &ICAPClient{
+		reqModURL:  cfg.ReqModURL,
+		respModURL: cfg.RespModURL,
+		pool:       make(chan *icapConn, maxPool),
+		log:        logger,
+		maxPool:    maxPool,
+	}
+}
+
+// ReqMod sends an HTTP request to the ICAP REQMOD service for inspection.
+// Returns the (possibly modified) request, or the original if ICAP returns 204.
+// Returns nil, nil if REQMOD is not configured.
+func (c *ICAPClient) ReqMod(req *http.Request) (*http.Request, error) {
+	if c.reqModURL == nil {
+		return req, nil
+	}
+
+	var reqBuf bytes.Buffer
+	if err := req.Write(&reqBuf); err != nil {
+		return nil, fmt.Errorf("serialize request: %w", err)
+	}
+
+	respBody, err := c.send("REQMOD", c.reqModURL, reqBuf.Bytes(), nil)
+	if err != nil {
+		return nil, err
+	}
+
+	if respBody == nil {
+		return req, nil
+	}
+
+	modified, err := http.ReadRequest(bufio.NewReader(bytes.NewReader(respBody)))
+	if err != nil {
+		return nil, fmt.Errorf("parse ICAP modified request: %w", err)
+	}
+	return modified, nil
+}
+
+// RespMod sends an HTTP response to the ICAP RESPMOD service for inspection.
+// Returns the (possibly modified) response, or the original if ICAP returns 204.
+// Returns nil, nil if RESPMOD is not configured.
+func (c *ICAPClient) RespMod(req *http.Request, resp *http.Response) (*http.Response, error) {
+	if c.respModURL == nil {
+		return resp, nil
+	}
+
+	var reqBuf bytes.Buffer
+	if err := req.Write(&reqBuf); err != nil {
+		return nil, fmt.Errorf("serialize request: %w", err)
+	}
+
+	var respBuf bytes.Buffer
+	if err := resp.Write(&respBuf); err != nil {
+		return nil, fmt.Errorf("serialize response: %w", err)
+	}
+
+	respBody, err := c.send("RESPMOD", c.respModURL, reqBuf.Bytes(), respBuf.Bytes())
+	if err != nil {
+		return nil, err
+	}
+
+	if respBody == nil {
+		// 204 No Content: ICAP server didn't modify the response.
+		// Reconstruct from the buffered copy since resp.Body was consumed by Write.
+		reconstructed, err := http.ReadResponse(bufio.NewReader(bytes.NewReader(respBuf.Bytes())), req)
+		if err != nil {
+			return nil, fmt.Errorf("reconstruct response after ICAP 204: %w", err)
+		}
+		return reconstructed, nil
+	}
+
+	modified, err := http.ReadResponse(bufio.NewReader(bytes.NewReader(respBody)), req)
+	if err != nil {
+		return nil, fmt.Errorf("parse ICAP modified response: %w", err)
+	}
+	return modified, nil
+}
+
+// Close drains and closes all pooled connections.
+func (c *ICAPClient) Close() {
+	close(c.pool)
+	for ic := range c.pool {
+		if err := ic.conn.Close(); err != nil {
+			c.log.Debugf("close ICAP connection: %v", err)
+		}
+	}
+}
+
+// send executes an ICAP request and returns the encapsulated body from the response.
+// Returns nil body for 204 No Content (no modification).
+// Retries once on stale pooled connection (EOF on read).
+func (c *ICAPClient) send(method string, serviceURL *url.URL, reqData, respData []byte) ([]byte, error) {
+	statusCode, headers, body, err := c.trySend(method, serviceURL, reqData, respData)
+	if err != nil && isStaleConnErr(err) {
+		// Retry once with a fresh connection (stale pool entry).
+		c.log.Debugf("ICAP %s: retrying after stale connection: %v", method, err)
+		statusCode, headers, body, err = c.trySend(method, serviceURL, reqData, respData)
+	}
+	if err != nil {
+		return nil, err
+	}
+
+	switch statusCode {
+	case 204:
+		return nil, nil
+	case 200:
+		return body, nil
+	default:
+		c.log.Debugf("ICAP %s returned status %d, headers: %v", method, statusCode, headers)
+		return nil, fmt.Errorf("ICAP %s: status %d", method, statusCode)
+	}
+}
+
+func (c *ICAPClient) trySend(method string, serviceURL *url.URL, reqData, respData []byte) (int, textproto.MIMEHeader, []byte, error) {
+	ic, err := c.getConn(serviceURL)
+	if err != nil {
+		return 0, nil, nil, fmt.Errorf("get ICAP connection: %w", err)
+	}
+
+	if err := c.writeRequest(ic, method, serviceURL, reqData, respData); err != nil {
+		if closeErr := ic.conn.Close(); closeErr != nil {
+			c.log.Debugf("close ICAP conn after write error: %v", closeErr)
+		}
+		return 0, nil, nil, fmt.Errorf("write ICAP %s: %w", method, err)
+	}
+
+	statusCode, headers, body, err := c.readResponse(ic)
+	if err != nil {
+		if closeErr := ic.conn.Close(); closeErr != nil {
+			c.log.Debugf("close ICAP conn after read error: %v", closeErr)
+		}
+		return 0, nil, nil, fmt.Errorf("read ICAP response: %w", err)
+	}
+
+	c.putConn(ic)
+	return statusCode, headers, body, nil
+}
+
+func isStaleConnErr(err error) bool {
+	if err == nil {
+		return false
+	}
+	s := err.Error()
+	return strings.Contains(s, "EOF") || strings.Contains(s, "broken pipe") || strings.Contains(s, "connection reset")
+}
+
+func (c *ICAPClient) writeRequest(ic *icapConn, method string, serviceURL *url.URL, reqData, respData []byte) error {
+	if err := ic.conn.SetWriteDeadline(time.Now().Add(icapRWTimeout)); err != nil {
+		return fmt.Errorf("set write deadline: %w", err)
+	}
+
+	// For RESPMOD, split the serialized HTTP response into headers and body.
+	// The body must be sent chunked per RFC 3507.
+	var respHdr, respBody []byte
+	if respData != nil {
+		if idx := bytes.Index(respData, []byte("\r\n\r\n")); idx >= 0 {
+			respHdr = respData[:idx+4] // include the \r\n\r\n separator
+			respBody = respData[idx+4:]
+		} else {
+			respHdr = respData
+		}
+	}
+
+	var buf bytes.Buffer
+
+	// Request line
+	fmt.Fprintf(&buf, "%s %s %s\r\n", method, serviceURL.String(), icapVersion)
+
+	// Headers
+	host := serviceURL.Host
+	fmt.Fprintf(&buf, "Host: %s\r\n", host)
+	fmt.Fprintf(&buf, "Connection: keep-alive\r\n")
+	fmt.Fprintf(&buf, "Allow: 204\r\n")
+
+	// Build Encapsulated header
+	offset := 0
+	var encapParts []string
+	if reqData != nil {
+		encapParts = append(encapParts, fmt.Sprintf("req-hdr=%d", offset))
+		offset += len(reqData)
+	}
+	if respHdr != nil {
+		encapParts = append(encapParts, fmt.Sprintf("res-hdr=%d", offset))
+		offset += len(respHdr)
+	}
+	if len(respBody) > 0 {
+		encapParts = append(encapParts, fmt.Sprintf("res-body=%d", offset))
+	} else {
+		encapParts = append(encapParts, fmt.Sprintf("null-body=%d", offset))
+	}
+	fmt.Fprintf(&buf, "Encapsulated: %s\r\n", strings.Join(encapParts, ", "))
+	fmt.Fprintf(&buf, "\r\n")
+
+	// Encapsulated sections
+	if reqData != nil {
+		buf.Write(reqData)
+	}
+	if respHdr != nil {
+		buf.Write(respHdr)
+	}
+	// Body in chunked encoding (only when there is an actual body section).
+	// Per RFC 3507 Section 4.4.1, null-body must not include any entity data.
+	if len(respBody) > 0 {
+		fmt.Fprintf(&buf, "%x\r\n", len(respBody))
+		buf.Write(respBody)
+		buf.WriteString("\r\n")
+		buf.WriteString("0\r\n\r\n")
+	}
+
+	_, err := ic.conn.Write(buf.Bytes())
+	return err
+}
+
+func (c *ICAPClient) readResponse(ic *icapConn) (int, textproto.MIMEHeader, []byte, error) {
+	if err := ic.conn.SetReadDeadline(time.Now().Add(icapRWTimeout)); err != nil {
+		return 0, nil, nil, fmt.Errorf("set read deadline: %w", err)
+	}
+
+	tp := textproto.NewReader(ic.reader)
+
+	// Status line: "ICAP/1.0 200 OK"
+	statusLine, err := tp.ReadLine()
+	if err != nil {
+		return 0, nil, nil, fmt.Errorf("read status line: %w", err)
+	}
+
+	statusCode, err := parseICAPStatus(statusLine)
+	if err != nil {
+		return 0, nil, nil, err
+	}
+
+	// Headers
+	headers, err := tp.ReadMIMEHeader()
+	if err != nil {
+		return statusCode, nil, nil, fmt.Errorf("read ICAP headers: %w", err)
+	}
+
+	if statusCode == 204 {
+		return statusCode, headers, nil, nil
+	}
+
+	// Read encapsulated body based on Encapsulated header
+	body, err := c.readEncapsulatedBody(ic.reader, headers)
+	if err != nil {
+		return statusCode, headers, nil, fmt.Errorf("read encapsulated body: %w", err)
+	}
+
+	return statusCode, headers, body, nil
+}
+
+func (c *ICAPClient) readEncapsulatedBody(r *bufio.Reader, headers textproto.MIMEHeader) ([]byte, error) {
+	encap := headers.Get("Encapsulated")
+	if encap == "" {
+		return nil, nil
+	}
+
+	// Find the body offset from the Encapsulated header.
+	// The last section with a non-zero offset is the body.
+	// Read everything from the reader as the encapsulated content.
+	var totalSize int
+	parts := strings.Split(encap, ",")
+	for _, part := range parts {
+		part = strings.TrimSpace(part)
+		eqIdx := strings.Index(part, "=")
+		if eqIdx < 0 {
+			continue
+		}
+		offset, err := strconv.Atoi(strings.TrimSpace(part[eqIdx+1:]))
+		if err != nil {
+			continue
+		}
+		if offset > totalSize {
+			totalSize = offset
+		}
+	}
+
+	// Read all available encapsulated data (headers + body)
+	// The body section uses chunked encoding per RFC 3507
+	var buf bytes.Buffer
+	if totalSize > 0 {
+		// Read the header sections (everything before the body offset)
+		headerBytes := make([]byte, totalSize)
+		if _, err := io.ReadFull(r, headerBytes); err != nil {
+			return nil, fmt.Errorf("read encapsulated headers: %w", err)
+		}
+		buf.Write(headerBytes)
+	}
+
+	// Read chunked body
+	chunked := newChunkedReader(r)
+	body, err := io.ReadAll(io.LimitReader(chunked, icapMaxRespSize))
+	if err != nil {
+		return nil, fmt.Errorf("read chunked body: %w", err)
+	}
+	buf.Write(body)
+
+	return buf.Bytes(), nil
+}
+
+func (c *ICAPClient) getConn(serviceURL *url.URL) (*icapConn, error) {
+	// Try to get a pooled connection
+	for {
+		select {
+		case ic := <-c.pool:
+			if time.Since(ic.lastUse) > icapIdleTimeout {
+				if err := ic.conn.Close(); err != nil {
+					c.log.Debugf("close idle ICAP connection: %v", err)
+				}
+				continue
+			}
+			return ic, nil
+		default:
+			return c.dialConn(serviceURL)
+		}
+	}
+}
+
+func (c *ICAPClient) putConn(ic *icapConn) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	ic.lastUse = time.Now()
+	select {
+	case c.pool <- ic:
+	default:
+		// Pool full, close connection.
+		if err := ic.conn.Close(); err != nil {
+			c.log.Debugf("close excess ICAP connection: %v", err)
+		}
+	}
+}
+
+func (c *ICAPClient) dialConn(serviceURL *url.URL) (*icapConn, error) {
+	host := serviceURL.Host
+	if _, _, err := net.SplitHostPort(host); err != nil {
+		host = net.JoinHostPort(host, icapDefaultPort)
+	}
+
+	conn, err := net.DialTimeout("tcp", host, icapConnTimeout)
+	if err != nil {
+		return nil, fmt.Errorf("dial ICAP %s: %w", host, err)
+	}
+
+	return &icapConn{
+		conn:    conn,
+		reader:  bufio.NewReader(conn),
+		lastUse: time.Now(),
+	}, nil
+}
+
+func parseICAPStatus(line string) (int, error) {
+	// "ICAP/1.0 200 OK"
+	parts := strings.SplitN(line, " ", 3)
+	if len(parts) < 2 {
+		return 0, fmt.Errorf("malformed ICAP status line: %q", line)
+	}
+	code, err := strconv.Atoi(parts[1])
+	if err != nil {
+		return 0, fmt.Errorf("parse ICAP status code %q: %w", parts[1], err)
+	}
+	return code, nil
+}
+
+// chunkedReader reads ICAP chunked encoding (same as HTTP chunked, terminated by "0\r\n\r\n").
+type chunkedReader struct {
+	r         *bufio.Reader
+	remaining int
+	done      bool
+}
+
+func newChunkedReader(r *bufio.Reader) *chunkedReader {
+	return &chunkedReader{r: r}
+}
+
+func (cr *chunkedReader) Read(p []byte) (int, error) {
+	if cr.done {
+		return 0, io.EOF
+	}
+
+	if cr.remaining == 0 {
+		// Read chunk size line
+		line, err := cr.r.ReadString('\n')
+		if err != nil {
+			return 0, err
+		}
+		line = strings.TrimSpace(line)
+
+		// Strip any chunk extensions
+		if idx := strings.Index(line, ";"); idx >= 0 {
+			line = line[:idx]
+		}
+
+		size, err := strconv.ParseInt(line, 16, 64)
+		if err != nil {
+			return 0, fmt.Errorf("parse chunk size %q: %w", line, err)
+		}
+
+		if size == 0 {
+			cr.done = true
+			// Consume trailing \r\n
+			_, _ = cr.r.ReadString('\n')
+			return 0, io.EOF
+		}
+
+		if size < 0 || size > icapMaxRespSize {
+			return 0, fmt.Errorf("chunk size %d out of range (max %d)", size, icapMaxRespSize)
+		}
+
+		cr.remaining = int(size)
+	}
+
+	toRead := len(p)
+	if toRead > cr.remaining {
+		toRead = cr.remaining
+	}
+
+	n, err := cr.r.Read(p[:toRead])
+	cr.remaining -= n
+
+	if cr.remaining == 0 {
+		// Consume chunk-terminating \r\n
+		_, _ = cr.r.ReadString('\n')
+	}
+
+	return n, err
+}

client/inspect/listener.go

@@ -0,0 +1,21 @@
+//go:build !linux
+
+package inspect
+
+import (
+	"fmt"
+	"net"
+	"net/netip"
+
+	log "github.com/sirupsen/logrus"
+)
+
+// newTPROXYListener is not supported on non-Linux platforms.
+func newTPROXYListener(_ *log.Entry, addr netip.AddrPort, _ netip.Prefix) (net.Listener, error) {
+	return nil, fmt.Errorf("TPROXY listener not supported on this platform (requested %s)", addr)
+}
+
+// getOriginalDst is not supported on non-Linux platforms.
+func getOriginalDst(_ net.Conn) (netip.AddrPort, error) {
+	return netip.AddrPort{}, fmt.Errorf("SO_ORIGINAL_DST not supported on this platform")
+}

client/inspect/listener_linux.go

@@ -0,0 +1,89 @@
+package inspect
+
+import (
+	"fmt"
+	"net"
+	"net/netip"
+	"unsafe"
+
+	log "github.com/sirupsen/logrus"
+	"golang.org/x/sys/unix"
+)
+
+// newTPROXYListener creates a TCP listener for the transparent proxy.
+// After nftables REDIRECT, accepted connections have LocalAddr = WG_IP:proxy_port.
+// The original destination is retrieved via getsockopt(SO_ORIGINAL_DST).
+func newTPROXYListener(logger *log.Entry, addr netip.AddrPort, _ netip.Prefix) (net.Listener, error) {
+	ln, err := net.Listen("tcp", addr.String())
+	if err != nil {
+		return nil, fmt.Errorf("listen on %s: %w", addr, err)
+	}
+
+	logger.Infof("inspect: listener started on %s", ln.Addr())
+	return ln, nil
+}
+
+// getOriginalDst reads the original destination from conntrack via SO_ORIGINAL_DST.
+// This is set by the kernel when the connection was REDIRECT'd/DNAT'd.
+// Tries IPv4 first, then falls back to IPv6 (IP6T_SO_ORIGINAL_DST).
+func getOriginalDst(conn net.Conn) (netip.AddrPort, error) {
+	tc, ok := conn.(*net.TCPConn)
+	if !ok {
+		return netip.AddrPort{}, fmt.Errorf("not a TCPConn")
+	}
+
+	raw, err := tc.SyscallConn()
+	if err != nil {
+		return netip.AddrPort{}, fmt.Errorf("get syscall conn: %w", err)
+	}
+
+	var origDst netip.AddrPort
+	var sockErr error
+	if err := raw.Control(func(fd uintptr) {
+		// Try IPv4 first (SO_ORIGINAL_DST = 80)
+		var sa4 unix.RawSockaddrInet4
+		sa4Len := uint32(unsafe.Sizeof(sa4))
+		_, _, errno := unix.Syscall6(
+			unix.SYS_GETSOCKOPT,
+			fd,
+			unix.SOL_IP,
+			80, // SO_ORIGINAL_DST
+			uintptr(unsafe.Pointer(&sa4)),
+			uintptr(unsafe.Pointer(&sa4Len)),
+			0,
+		)
+		if errno == 0 {
+			addr := netip.AddrFrom4(sa4.Addr)
+			port := uint16(sa4.Port>>8) | uint16(sa4.Port<<8)
+			origDst = netip.AddrPortFrom(addr.Unmap(), port)
+			return
+		}
+
+		// Fall back to IPv6 (IP6T_SO_ORIGINAL_DST = 80 on SOL_IPV6)
+		var sa6 unix.RawSockaddrInet6
+		sa6Len := uint32(unsafe.Sizeof(sa6))
+		_, _, errno = unix.Syscall6(
+			unix.SYS_GETSOCKOPT,
+			fd,
+			unix.SOL_IPV6,
+			80, // IP6T_SO_ORIGINAL_DST
+			uintptr(unsafe.Pointer(&sa6)),
+			uintptr(unsafe.Pointer(&sa6Len)),
+			0,
+		)
+		if errno != 0 {
+			sockErr = fmt.Errorf("getsockopt SO_ORIGINAL_DST (v4 and v6): %w", errno)
+			return
+		}
+		addr := netip.AddrFrom16(sa6.Addr)
+		port := uint16(sa6.Port>>8) | uint16(sa6.Port<<8)
+		origDst = netip.AddrPortFrom(addr.Unmap(), port)
+	}); err != nil {
+		return netip.AddrPort{}, fmt.Errorf("control raw conn: %w", err)
+	}
+	if sockErr != nil {
+		return netip.AddrPort{}, sockErr
+	}
+
+	return origDst, nil
+}

client/inspect/mitm.go

@@ -0,0 +1,200 @@
+package inspect
+
+import (
+	"crypto"
+	"crypto/ecdsa"
+	"crypto/elliptic"
+	"crypto/rand"
+	"crypto/tls"
+	"crypto/x509"
+	"crypto/x509/pkix"
+	"fmt"
+	"math/big"
+	mrand "math/rand/v2"
+	"sync"
+	"time"
+)
+
+const (
+	// certCacheSize is the maximum number of cached leaf certificates.
+	certCacheSize = 1024
+	// certTTL is how long generated certificates remain valid.
+	certTTL = 24 * time.Hour
+)
+
+// certCache is a bounded LRU cache for generated TLS certificates.
+type certCache struct {
+	mu      sync.Mutex
+	entries map[string]*certEntry
+	// order tracks LRU eviction, most recent at end.
+	order   []string
+	maxSize int
+}
+
+type certEntry struct {
+	cert      *tls.Certificate
+	expiresAt time.Time
+}
+
+func newCertCache(maxSize int) *certCache {
+	return &certCache{
+		entries: make(map[string]*certEntry, maxSize),
+		order:   make([]string, 0, maxSize),
+		maxSize: maxSize,
+	}
+}
+
+func (c *certCache) get(hostname string) (*tls.Certificate, bool) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	entry, ok := c.entries[hostname]
+	if !ok {
+		return nil, false
+	}
+
+	if time.Now().After(entry.expiresAt) {
+		c.removeLocked(hostname)
+		return nil, false
+	}
+
+	// Move to end (most recently used)
+	c.touchLocked(hostname)
+	return entry.cert, true
+}
+
+func (c *certCache) put(hostname string, cert *tls.Certificate) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	// Jitter the TTL by +/- 20% to prevent thundering herd on expiry.
+	jitter := time.Duration(float64(certTTL) * (0.8 + 0.4*mrand.Float64()))
+
+	if _, exists := c.entries[hostname]; exists {
+		c.entries[hostname] = &certEntry{
+			cert:      cert,
+			expiresAt: time.Now().Add(jitter),
+		}
+		c.touchLocked(hostname)
+		return
+	}
+
+	// Evict oldest if at capacity
+	for len(c.entries) >= c.maxSize && len(c.order) > 0 {
+		c.removeLocked(c.order[0])
+	}
+
+	c.entries[hostname] = &certEntry{
+		cert:      cert,
+		expiresAt: time.Now().Add(jitter),
+	}
+	c.order = append(c.order, hostname)
+}
+
+func (c *certCache) touchLocked(hostname string) {
+	for i, h := range c.order {
+		if h == hostname {
+			c.order = append(c.order[:i], c.order[i+1:]...)
+			c.order = append(c.order, hostname)
+			return
+		}
+	}
+}
+
+func (c *certCache) removeLocked(hostname string) {
+	delete(c.entries, hostname)
+	for i, h := range c.order {
+		if h == hostname {
+			c.order = append(c.order[:i], c.order[i+1:]...)
+			return
+		}
+	}
+}
+
+// CertProvider generates TLS certificates on the fly, signed by a CA.
+// Generated certificates are cached in an LRU cache.
+type CertProvider struct {
+	ca    *x509.Certificate
+	caKey crypto.PrivateKey
+	cache *certCache
+}
+
+// NewCertProvider creates a certificate provider using the given CA.
+func NewCertProvider(ca *x509.Certificate, caKey crypto.PrivateKey) *CertProvider {
+	return &CertProvider{
+		ca:    ca,
+		caKey: caKey,
+		cache: newCertCache(certCacheSize),
+	}
+}
+
+// GetCertificate returns a TLS certificate for the given hostname,
+// generating and caching one if necessary.
+func (p *CertProvider) GetCertificate(hostname string) (*tls.Certificate, error) {
+	if cert, ok := p.cache.get(hostname); ok {
+		return cert, nil
+	}
+
+	cert, err := p.generateCert(hostname)
+	if err != nil {
+		return nil, fmt.Errorf("generate cert for %s: %w", hostname, err)
+	}
+
+	p.cache.put(hostname, cert)
+	return cert, nil
+}
+
+// GetTLSConfig returns a tls.Config that dynamically provides certificates
+// for any hostname using the MITM CA.
+func (p *CertProvider) GetTLSConfig() *tls.Config {
+	return &tls.Config{
+		GetCertificate: func(hello *tls.ClientHelloInfo) (*tls.Certificate, error) {
+			return p.GetCertificate(hello.ServerName)
+		},
+		NextProtos: []string{"h2", "http/1.1"},
+		MinVersion: tls.VersionTLS12,
+	}
+}
+
+func (p *CertProvider) generateCert(hostname string) (*tls.Certificate, error) {
+	serialNumber, err := rand.Int(rand.Reader, new(big.Int).Lsh(big.NewInt(1), 128))
+	if err != nil {
+		return nil, fmt.Errorf("generate serial number: %w", err)
+	}
+
+	now := time.Now()
+	template := &x509.Certificate{
+		SerialNumber: serialNumber,
+		Subject: pkix.Name{
+			CommonName: hostname,
+		},
+		NotBefore: now.Add(-5 * time.Minute),
+		NotAfter:  now.Add(certTTL),
+		KeyUsage:  x509.KeyUsageDigitalSignature | x509.KeyUsageKeyEncipherment,
+		ExtKeyUsage: []x509.ExtKeyUsage{
+			x509.ExtKeyUsageServerAuth,
+		},
+		DNSNames: []string{hostname},
+	}
+
+	leafKey, err := ecdsa.GenerateKey(elliptic.P256(), rand.Reader)
+	if err != nil {
+		return nil, fmt.Errorf("generate leaf key: %w", err)
+	}
+
+	certDER, err := x509.CreateCertificate(rand.Reader, template, p.ca, &leafKey.PublicKey, p.caKey)
+	if err != nil {
+		return nil, fmt.Errorf("sign leaf certificate: %w", err)
+	}
+
+	leafCert, err := x509.ParseCertificate(certDER)
+	if err != nil {
+		return nil, fmt.Errorf("parse generated certificate: %w", err)
+	}
+
+	return &tls.Certificate{
+		Certificate: [][]byte{certDER, p.ca.Raw},
+		PrivateKey:  leafKey,
+		Leaf:        leafCert,
+	}, nil
+}

client/inspect/mitm_test.go

@@ -0,0 +1,133 @@
+package inspect
+
+import (
+	"crypto/ecdsa"
+	"crypto/elliptic"
+	"crypto/rand"
+	"crypto/tls"
+	"crypto/x509"
+	"crypto/x509/pkix"
+	"math/big"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func generateTestCA(t *testing.T) (*x509.Certificate, *ecdsa.PrivateKey) {
+	t.Helper()
+
+	key, err := ecdsa.GenerateKey(elliptic.P256(), rand.Reader)
+	require.NoError(t, err)
+
+	template := &x509.Certificate{
+		SerialNumber: big.NewInt(1),
+		Subject: pkix.Name{
+			CommonName: "Test CA",
+		},
+		NotBefore:             time.Now().Add(-time.Hour),
+		NotAfter:              time.Now().Add(24 * time.Hour),
+		KeyUsage:              x509.KeyUsageCertSign | x509.KeyUsageCRLSign,
+		BasicConstraintsValid: true,
+		IsCA:                  true,
+	}
+
+	certDER, err := x509.CreateCertificate(rand.Reader, template, template, &key.PublicKey, key)
+	require.NoError(t, err)
+
+	cert, err := x509.ParseCertificate(certDER)
+	require.NoError(t, err)
+
+	return cert, key
+}
+
+func TestCertProvider_GetCertificate(t *testing.T) {
+	ca, caKey := generateTestCA(t)
+	provider := NewCertProvider(ca, caKey)
+
+	cert, err := provider.GetCertificate("example.com")
+	require.NoError(t, err)
+	require.NotNil(t, cert)
+
+	// Verify the leaf certificate
+	assert.Equal(t, "example.com", cert.Leaf.Subject.CommonName)
+	assert.Contains(t, cert.Leaf.DNSNames, "example.com")
+
+	// Verify chain: leaf + CA
+	assert.Len(t, cert.Certificate, 2)
+
+	// Verify leaf is signed by our CA
+	pool := x509.NewCertPool()
+	pool.AddCert(ca)
+	_, err = cert.Leaf.Verify(x509.VerifyOptions{
+		Roots: pool,
+	})
+	require.NoError(t, err)
+}
+
+func TestCertProvider_CachesResults(t *testing.T) {
+	ca, caKey := generateTestCA(t)
+	provider := NewCertProvider(ca, caKey)
+
+	cert1, err := provider.GetCertificate("cached.example.com")
+	require.NoError(t, err)
+
+	cert2, err := provider.GetCertificate("cached.example.com")
+	require.NoError(t, err)
+
+	// Same pointer = cached
+	assert.Equal(t, cert1, cert2)
+}
+
+func TestCertProvider_DifferentHostsDifferentCerts(t *testing.T) {
+	ca, caKey := generateTestCA(t)
+	provider := NewCertProvider(ca, caKey)
+
+	cert1, err := provider.GetCertificate("a.example.com")
+	require.NoError(t, err)
+
+	cert2, err := provider.GetCertificate("b.example.com")
+	require.NoError(t, err)
+
+	assert.NotEqual(t, cert1.Leaf.SerialNumber, cert2.Leaf.SerialNumber)
+}
+
+func TestCertProvider_TLSConfigHandshake(t *testing.T) {
+	ca, caKey := generateTestCA(t)
+	provider := NewCertProvider(ca, caKey)
+
+	tlsConfig := provider.GetTLSConfig()
+	require.NotNil(t, tlsConfig)
+	require.NotNil(t, tlsConfig.GetCertificate)
+
+	// Simulate a ClientHelloInfo
+	hello := &tls.ClientHelloInfo{
+		ServerName: "handshake.example.com",
+	}
+
+	cert, err := tlsConfig.GetCertificate(hello)
+	require.NoError(t, err)
+	assert.Equal(t, "handshake.example.com", cert.Leaf.Subject.CommonName)
+}
+
+func TestCertCache_Eviction(t *testing.T) {
+	cache := newCertCache(3)
+
+	for i := range 5 {
+		hostname := string(rune('a'+i)) + ".example.com"
+		cache.put(hostname, &tls.Certificate{})
+	}
+
+	// Only 3 should remain (c, d, e - the most recent)
+	assert.Len(t, cache.entries, 3)
+
+	_, ok := cache.get("a.example.com")
+	assert.False(t, ok, "oldest entry should be evicted")
+
+	_, ok = cache.get("b.example.com")
+	assert.False(t, ok, "second oldest should be evicted")
+
+	_, ok = cache.get("e.example.com")
+	assert.True(t, ok, "newest entry should exist")
+}

client/inspect/peek.go

@@ -0,0 +1,109 @@
+package inspect
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"net"
+)
+
+// peekConn wraps a net.Conn with a buffer that allows reading ahead
+// without consuming data. Subsequent Read calls return the buffered
+// bytes first, then read from the underlying connection.
+type peekConn struct {
+	net.Conn
+	buf bytes.Buffer
+	// peeked holds the raw bytes that were peeked, available for replay.
+	peeked []byte
+}
+
+// newPeekConn wraps conn for peek-ahead reading.
+func newPeekConn(conn net.Conn) *peekConn {
+	return &peekConn{Conn: conn}
+}
+
+// Peek reads exactly n bytes from the connection without consuming them.
+// The peeked bytes are replayed on subsequent Read calls.
+// Peek may only be called once; calling it again returns an error.
+func (c *peekConn) Peek(n int) ([]byte, error) {
+	if c.peeked != nil {
+		return nil, fmt.Errorf("peek already called")
+	}
+
+	buf := make([]byte, n)
+	if _, err := io.ReadFull(c.Conn, buf); err != nil {
+		return nil, fmt.Errorf("peek %d bytes: %w", n, err)
+	}
+
+	c.peeked = buf
+	c.buf.Write(buf)
+
+	return buf, nil
+}
+
+// PeekAll reads up to n bytes, returning whatever is available.
+// Unlike Peek, it does not require exactly n bytes.
+func (c *peekConn) PeekAll(n int) ([]byte, error) {
+	if c.peeked != nil {
+		return nil, fmt.Errorf("peek already called")
+	}
+
+	buf := make([]byte, n)
+	nr, err := c.Conn.Read(buf)
+	if nr > 0 {
+		c.peeked = buf[:nr]
+		c.buf.Write(c.peeked)
+	}
+	if err != nil && nr == 0 {
+		return nil, fmt.Errorf("peek: %w", err)
+	}
+
+	return c.peeked, nil
+}
+
+// PeekMore extends the peeked buffer to at least n total bytes.
+// The buffer is reset and refilled with the extended data.
+// The returned slice is the internal peeked buffer; callers must not
+// retain references from prior Peek/PeekMore calls after calling this.
+func (c *peekConn) PeekMore(n int) ([]byte, error) {
+	if len(c.peeked) >= n {
+		return c.peeked[:n], nil
+	}
+
+	remaining := n - len(c.peeked)
+	extra := make([]byte, remaining)
+	if _, err := io.ReadFull(c.Conn, extra); err != nil {
+		return nil, fmt.Errorf("peek more %d bytes: %w", remaining, err)
+	}
+
+	// Pre-allocate to avoid reallocation detaching previously returned slices.
+	combined := make([]byte, 0, n)
+	combined = append(combined, c.peeked...)
+	combined = append(combined, extra...)
+	c.peeked = combined
+	c.buf.Reset()
+	c.buf.Write(c.peeked)
+
+	return c.peeked, nil
+}
+
+// Peeked returns the bytes that were peeked so far, or nil if Peek hasn't been called.
+func (c *peekConn) Peeked() []byte {
+	return c.peeked
+}
+
+// Read returns buffered peek data first, then reads from the underlying connection.
+func (c *peekConn) Read(p []byte) (int, error) {
+	if c.buf.Len() > 0 {
+		return c.buf.Read(p)
+	}
+	return c.Conn.Read(p)
+}
+
+// reader returns an io.Reader that replays buffered bytes then reads from conn.
+func (c *peekConn) reader() io.Reader {
+	if c.buf.Len() > 0 {
+		return io.MultiReader(&c.buf, c.Conn)
+	}
+	return c.Conn
+}

client/inspect/proxy.go

@@ -0,0 +1,482 @@
+package inspect
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"net"
+	"net/netip"
+	"sync"
+	"time"
+
+	log "github.com/sirupsen/logrus"
+)
+
+// ErrBlocked is returned when a connection is denied by proxy policy.
+var ErrBlocked = errors.New("connection blocked by proxy policy")
+
+const (
+	// headerReadTimeout is the deadline for reading the initial protocol header.
+	// Prevents slow loris attacks where a client opens a connection but sends data slowly.
+	headerReadTimeout = 10 * time.Second
+
+	// idleTimeout is the deadline for idle connections between HTTP requests.
+	idleTimeout = 120 * time.Second
+)
+
+// Proxy is the inspection engine for traffic passing through a NetBird
+// routing peer. It handles protocol detection, rule evaluation, MITM TLS
+// decryption, ICAP delegation, and external proxy forwarding.
+type Proxy struct {
+	config Config
+	rules  *RuleEngine
+	certs  *CertProvider
+	icap   *ICAPClient
+	// envoy is nil unless mode is ModeEnvoy.
+	envoy *envoyManager
+	// dialer is the outbound dialer (with SO_MARK cleared on Linux).
+	dialer net.Dialer
+	log    *log.Entry
+	// wgNetwork is the WG overlay prefix; dial targets inside it are blocked.
+	wgNetwork netip.Prefix
+	// localIPs reports the routing peer's own IPs; dial targets are blocked.
+	localIPs LocalIPChecker
+	// listener is the TPROXY/REDIRECT listener for kernel mode.
+	listener net.Listener
+
+	mu     sync.RWMutex
+	ctx    context.Context
+	cancel context.CancelFunc
+}
+
+// LocalIPChecker reports whether an IP belongs to the local machine.
+type LocalIPChecker interface {
+	IsLocalIP(netip.Addr) bool
+}
+
+// New creates a transparent proxy with the given configuration.
+func New(ctx context.Context, logger *log.Entry, config Config) (*Proxy, error) {
+	ctx, cancel := context.WithCancel(ctx)
+
+	p := &Proxy{
+		config:    config,
+		rules:     NewRuleEngine(logger, config.DefaultAction),
+		dialer:    newOutboundDialer(),
+		log:       logger,
+		wgNetwork: config.WGNetwork,
+		localIPs:  config.LocalIPChecker,
+		ctx:       ctx,
+		cancel:    cancel,
+	}
+
+	p.rules.UpdateRules(config.Rules, config.DefaultAction)
+
+	// Initialize MITM certificate provider
+	if config.TLS != nil {
+		p.certs = NewCertProvider(config.TLS.CA, config.TLS.CAKey)
+	}
+
+	// Initialize ICAP client
+	if config.ICAP != nil {
+		p.icap = NewICAPClient(logger, config.ICAP)
+	}
+
+	// Start envoy sidecar if configured
+	if config.Mode == ModeEnvoy {
+		envoyLog := logger.WithField("sidecar", "envoy")
+		em, err := startEnvoy(ctx, envoyLog, config)
+		if err != nil {
+			cancel()
+			return nil, fmt.Errorf("start envoy sidecar: %w", err)
+		}
+		p.envoy = em
+	}
+
+	// Start TPROXY listener for kernel mode
+	if config.ListenAddr.IsValid() {
+		ln, err := newTPROXYListener(logger, config.ListenAddr, netip.Prefix{})
+		if err != nil {
+			cancel()
+			return nil, fmt.Errorf("start TPROXY listener on %s: %w", config.ListenAddr, err)
+		}
+		p.listener = ln
+		go p.acceptLoop(ln)
+	}
+
+	return p, nil
+}
+
+// HandleTCP is the entry point for TCP connections from the userspace forwarder.
+// It determines the protocol (TLS or plaintext HTTP), evaluates rules,
+// and either blocks, passes through, inspects, or forwards to an external proxy.
+func (p *Proxy) HandleTCP(ctx context.Context, clientConn net.Conn, dst netip.AddrPort, src SourceInfo) error {
+	defer func() {
+		if err := clientConn.Close(); err != nil {
+			p.log.Debugf("close client conn: %v", err)
+		}
+	}()
+
+	p.mu.RLock()
+	mode := p.config.Mode
+	p.mu.RUnlock()
+
+	if mode == ModeExternal {
+		pconn := newPeekConn(clientConn)
+		return p.handleExternal(ctx, pconn, dst)
+	}
+
+	// Envoy and builtin modes both peek the protocol header for rule evaluation.
+	// Envoy mode forwards non-blocked traffic to envoy; builtin mode handles all locally.
+	// TLS blocks are handled by Go (instant close) since envoy can't cleanly RST a TLS connection.
+
+	// Built-in and envoy mode: peek 5 bytes (TLS record header size) to determine protocol.
+	// Set a read deadline to prevent slow loris attacks.
+	if err := clientConn.SetReadDeadline(time.Now().Add(headerReadTimeout)); err != nil {
+		return fmt.Errorf("set read deadline: %w", err)
+	}
+	pconn := newPeekConn(clientConn)
+	header, err := pconn.Peek(5)
+	if err != nil {
+		return fmt.Errorf("peek protocol header: %w", err)
+	}
+	if err := clientConn.SetReadDeadline(time.Time{}); err != nil {
+		return fmt.Errorf("clear read deadline: %w", err)
+	}
+
+	if isTLSHandshake(header[0]) {
+		return p.handleTLS(ctx, pconn, dst, src)
+	}
+
+	if isHTTPMethod(header) {
+		return p.handlePlainHTTP(ctx, pconn, dst, src)
+	}
+
+	// Not TLS and not HTTP: evaluate rules with ProtoOther.
+	// If no rule explicitly allows "other", this falls through to the default action.
+	action := p.rules.Evaluate(src.IP, "", dst.Addr(), dst.Port(), ProtoOther, "")
+	if action == ActionAllow {
+		remote, err := p.dialTCP(ctx, dst)
+		if err != nil {
+			return fmt.Errorf("dial for passthrough: %w", err)
+		}
+		defer func() {
+			if err := remote.Close(); err != nil {
+				p.log.Debugf("close remote conn: %v", err)
+			}
+		}()
+		return relay(ctx, pconn, remote)
+	}
+
+	p.log.Debugf("block: non-HTTP/TLS to %s (action=%s, first bytes: %x)", dst, action, header)
+	return ErrBlocked
+}
+
+// InspectTCP evaluates rules for a TCP connection and returns the result.
+// Unlike HandleTCP, it can return early for allow decisions, letting the caller
+// handle the relay (USP forwarder passthrough optimization).
+//
+// When InspectResult.PassthroughConn is non-nil, ownership transfers to the caller:
+// the caller must close the connection and relay traffic. The engine does not close it.
+//
+// When PassthroughConn is nil, the engine handled everything internally
+// (block, inspect/MITM, or plain HTTP inspection) and closed the connection.
+func (p *Proxy) InspectTCP(ctx context.Context, clientConn net.Conn, dst netip.AddrPort, src SourceInfo) (InspectResult, error) {
+	p.mu.RLock()
+	mode := p.config.Mode
+	envoy := p.envoy
+	p.mu.RUnlock()
+
+	// External mode: handle internally, engine owns the connection.
+	if mode == ModeExternal {
+		defer func() {
+			if err := clientConn.Close(); err != nil {
+				p.log.Debugf("close client conn: %v", err)
+			}
+		}()
+		pconn := newPeekConn(clientConn)
+		err := p.handleExternal(ctx, pconn, dst)
+		return InspectResult{Action: ActionAllow}, err
+	}
+
+	// Peek protocol header.
+	if err := clientConn.SetReadDeadline(time.Now().Add(headerReadTimeout)); err != nil {
+		clientConn.Close()
+		return InspectResult{}, fmt.Errorf("set read deadline: %w", err)
+	}
+	pconn := newPeekConn(clientConn)
+	header, err := pconn.Peek(5)
+	if err != nil {
+		clientConn.Close()
+		return InspectResult{}, fmt.Errorf("peek protocol header: %w", err)
+	}
+	if err := clientConn.SetReadDeadline(time.Time{}); err != nil {
+		clientConn.Close()
+		return InspectResult{}, fmt.Errorf("clear read deadline: %w", err)
+	}
+
+	// TLS: may return passthrough for allow.
+	if isTLSHandshake(header[0]) {
+		result, err := p.inspectTLS(ctx, pconn, dst, src)
+		if err != nil && result.PassthroughConn == nil {
+			clientConn.Close()
+			return result, err
+		}
+		// Envoy mode: forward allowed TLS to envoy instead of returning passthrough.
+		if result.PassthroughConn != nil && envoy != nil {
+			defer clientConn.Close()
+			envoyErr := p.forwardToEnvoy(ctx, pconn, dst, src, envoy)
+			return InspectResult{Action: ActionAllow}, envoyErr
+		}
+		return result, err
+	}
+
+	// Plain HTTP: in envoy mode, forward to envoy for L7 processing.
+	// In builtin mode, inspect per-request locally.
+	if isHTTPMethod(header) {
+		defer func() {
+			if err := clientConn.Close(); err != nil {
+				p.log.Debugf("close client conn: %v", err)
+			}
+		}()
+		if envoy != nil {
+			err := p.forwardToEnvoy(ctx, pconn, dst, src, envoy)
+			return InspectResult{Action: ActionAllow}, err
+		}
+		err := p.handlePlainHTTP(ctx, pconn, dst, src)
+		return InspectResult{Action: ActionInspect}, err
+	}
+
+	// Other protocol: evaluate rules.
+	action := p.rules.Evaluate(src.IP, "", dst.Addr(), dst.Port(), ProtoOther, "")
+	if action == ActionAllow {
+		// Envoy mode: forward to envoy.
+		if envoy != nil {
+			defer clientConn.Close()
+			err := p.forwardToEnvoy(ctx, pconn, dst, src, envoy)
+			return InspectResult{Action: ActionAllow}, err
+		}
+		return InspectResult{Action: ActionAllow, PassthroughConn: pconn}, nil
+	}
+
+	p.log.Debugf("block: non-HTTP/TLS to %s (action=%s, first bytes: %x)", dst, action, header)
+	clientConn.Close()
+	return InspectResult{Action: ActionBlock}, ErrBlocked
+}
+
+// HandleUDPPacket inspects a UDP packet for QUIC Initial packets.
+// Returns the action to take: ActionAllow to continue normal forwarding,
+// ActionBlock to drop the packet.
+// Non-QUIC packets always return ActionAllow.
+func (p *Proxy) HandleUDPPacket(data []byte, dst netip.AddrPort, src SourceInfo) Action {
+	if len(data) < 5 {
+		return ActionAllow
+	}
+
+	// Check for QUIC Long Header
+	if data[0]&0x80 == 0 {
+		return ActionAllow
+	}
+
+	sni, err := ExtractQUICSNI(data)
+	if err != nil {
+		// Can't parse QUIC, allow through (could be non-QUIC UDP)
+		p.log.Tracef("QUIC SNI extraction failed for %s: %v", dst, err)
+		return ActionAllow
+	}
+
+	if sni == "" {
+		return ActionAllow
+	}
+
+	action := p.rules.Evaluate(src.IP, sni, dst.Addr(), dst.Port(), ProtoH3, "")
+
+	if action == ActionBlock {
+		p.log.Debugf("block: QUIC to %s (SNI=%s)", dst, sni.PunycodeString())
+		return ActionBlock
+	}
+
+	// QUIC can't be MITMed, treat Inspect as Allow
+	if action == ActionInspect {
+		p.log.Debugf("allow: QUIC to %s (SNI=%s), MITM not supported for QUIC", dst, sni.PunycodeString())
+	} else {
+		p.log.Tracef("allow: QUIC to %s (SNI=%s)", dst, sni.PunycodeString())
+	}
+
+	return ActionAllow
+}
+
+// handlePlainHTTP handles plaintext HTTP connections.
+func (p *Proxy) handlePlainHTTP(ctx context.Context, pconn *peekConn, dst netip.AddrPort, src SourceInfo) error {
+	remote, err := p.dialTCP(ctx, dst)
+	if err != nil {
+		return fmt.Errorf("dial %s: %w", dst, err)
+	}
+	defer func() {
+		if err := remote.Close(); err != nil {
+			p.log.Debugf("close remote for %s: %v", dst, err)
+		}
+	}()
+
+	// For plaintext HTTP, always inspect (we can see the traffic)
+	return p.inspectHTTP(ctx, pconn, remote, dst, "", src, "http/1.1")
+}
+
+// UpdateConfig replaces the inspection engine configuration at runtime.
+func (p *Proxy) UpdateConfig(config Config) {
+	p.log.Debugf("config update: mode=%s rules=%d default=%s has_tls=%v has_icap=%v",
+		config.Mode, len(config.Rules), config.DefaultAction, config.TLS != nil, config.ICAP != nil)
+
+	p.mu.Lock()
+
+	p.config = config
+	p.rules.UpdateRules(config.Rules, config.DefaultAction)
+
+	// Update MITM provider
+	if config.TLS != nil {
+		p.certs = NewCertProvider(config.TLS.CA, config.TLS.CAKey)
+	} else {
+		p.certs = nil
+	}
+
+	// Swap ICAP client under lock, close the old one outside to avoid blocking.
+	var oldICAP *ICAPClient
+	if config.ICAP != nil {
+		oldICAP = p.icap
+		p.icap = NewICAPClient(p.log, config.ICAP)
+	} else {
+		oldICAP = p.icap
+		p.icap = nil
+	}
+
+	// If switching away from envoy mode, clear and stop the old envoy.
+	var oldEnvoy *envoyManager
+	if config.Mode != ModeEnvoy && p.envoy != nil {
+		oldEnvoy = p.envoy
+		p.envoy = nil
+	}
+
+	envoy := p.envoy
+
+	p.mu.Unlock()
+
+	if oldICAP != nil {
+		oldICAP.Close()
+	}
+
+	if oldEnvoy != nil {
+		oldEnvoy.Stop()
+	}
+
+	// Reload envoy config if still in envoy mode.
+	if envoy != nil && config.Mode == ModeEnvoy {
+		if err := envoy.Reload(config); err != nil {
+			p.log.Errorf("inspect: envoy config reload: %v", err)
+		}
+	}
+}
+
+// Mode returns the current proxy operating mode.
+func (p *Proxy) Mode() ProxyMode {
+	p.mu.RLock()
+	defer p.mu.RUnlock()
+	return p.config.Mode
+}
+
+// ListenPort returns the port to use for kernel-mode nftables REDIRECT.
+// For builtin mode: the TPROXY listener port.
+// For envoy mode: the envoy listener port (nftables redirects directly to envoy).
+// Returns 0 if no listener is active.
+func (p *Proxy) ListenPort() uint16 {
+	p.mu.RLock()
+	envoy := p.envoy
+	p.mu.RUnlock()
+
+	if envoy != nil {
+		return envoy.listenPort
+	}
+	if p.listener == nil {
+		return 0
+	}
+	tcpAddr, ok := p.listener.Addr().(*net.TCPAddr)
+	if !ok {
+		return 0
+	}
+	return uint16(tcpAddr.Port)
+}
+
+// Close shuts down the proxy and releases resources.
+func (p *Proxy) Close() error {
+	p.cancel()
+
+	p.mu.Lock()
+	envoy := p.envoy
+	p.envoy = nil
+	icap := p.icap
+	p.icap = nil
+	p.mu.Unlock()
+
+	if envoy != nil {
+		envoy.Stop()
+	}
+
+	if p.listener != nil {
+		if err := p.listener.Close(); err != nil {
+			p.log.Debugf("close TPROXY listener: %v", err)
+		}
+	}
+
+	if icap != nil {
+		icap.Close()
+	}
+
+	return nil
+}
+
+// acceptLoop accepts connections from the redirected listener (kernel mode).
+// Connections arrive via nftables REDIRECT; original destination is read from conntrack.
+func (p *Proxy) acceptLoop(ln net.Listener) {
+	for {
+		conn, err := ln.Accept()
+		if err != nil {
+			if p.ctx.Err() != nil {
+				return
+			}
+			p.log.Debugf("accept error: %v", err)
+			continue
+		}
+
+		go func() {
+			// Read original destination from conntrack (SO_ORIGINAL_DST).
+			// nftables REDIRECT changes dst to the local WG IP:proxy_port,
+			// but conntrack preserves the real destination.
+			dstAddr, err := getOriginalDst(conn)
+			if err != nil {
+				p.log.Debugf("get original dst: %v", err)
+				if closeErr := conn.Close(); closeErr != nil {
+					p.log.Debugf("close conn: %v", closeErr)
+				}
+				return
+			}
+
+			p.log.Tracef("accepted: %s -> %s (original dst %s)",
+				conn.RemoteAddr(), conn.LocalAddr(), dstAddr)
+
+			srcAddr, err := netip.ParseAddrPort(conn.RemoteAddr().String())
+			if err != nil {
+				p.log.Debugf("parse source: %v", err)
+				if closeErr := conn.Close(); closeErr != nil {
+					p.log.Debugf("close conn: %v", closeErr)
+				}
+				return
+			}
+
+			src := SourceInfo{
+				IP: srcAddr.Addr().Unmap(),
+			}
+
+			if err := p.HandleTCP(p.ctx, conn, dstAddr, src); err != nil && !errors.Is(err, ErrBlocked) {
+				p.log.Debugf("connection to %s: %v", dstAddr, err)
+			}
+		}()
+	}
+}

client/inspect/quic.go

@@ -0,0 +1,388 @@
+package inspect
+
+import (
+	"crypto/aes"
+	"crypto/cipher"
+	"crypto/sha256"
+	"encoding/binary"
+	"fmt"
+	"io"
+
+	"golang.org/x/crypto/hkdf"
+
+	"github.com/netbirdio/netbird/shared/management/domain"
+)
+
+// QUIC version constants
+const (
+	quicV1Version uint32 = 0x00000001
+	quicV2Version uint32 = 0x6b3343cf
+)
+
+// quicV1Salt is the initial salt for QUIC v1 (RFC 9001 Section 5.2).
+var quicV1Salt = []byte{
+	0x38, 0x76, 0x2c, 0xf7, 0xf5, 0x59, 0x34, 0xb3,
+	0x4d, 0x17, 0x9a, 0xe6, 0xa4, 0xc8, 0x0c, 0xad,
+	0xcc, 0xbb, 0x7f, 0x0a,
+}
+
+// quicV2Salt is the initial salt for QUIC v2 (RFC 9369).
+var quicV2Salt = []byte{
+	0x0d, 0xed, 0xe3, 0xde, 0xf7, 0x00, 0xa6, 0xdb,
+	0x81, 0x93, 0x81, 0xbe, 0x6e, 0x26, 0x9d, 0xcb,
+	0xf9, 0xbd, 0x2e, 0xd9,
+}
+
+// ExtractQUICSNI extracts the SNI from a QUIC Initial packet.
+// The Initial packet's encryption uses well-known keys derived from the
+// Destination Connection ID, so any observer can decrypt it (by design).
+func ExtractQUICSNI(data []byte) (domain.Domain, error) {
+	if len(data) < 5 {
+		return "", fmt.Errorf("packet too short")
+	}
+
+	// Check for QUIC Long Header (form bit set)
+	if data[0]&0x80 == 0 {
+		return "", fmt.Errorf("not a QUIC long header packet")
+	}
+
+	// Version
+	version := binary.BigEndian.Uint32(data[1:5])
+
+	var salt []byte
+	var initialLabel, keyLabel, ivLabel, hpLabel string
+
+	switch version {
+	case quicV1Version:
+		salt = quicV1Salt
+		initialLabel = "client in"
+		keyLabel = "quic key"
+		ivLabel = "quic iv"
+		hpLabel = "quic hp"
+	case quicV2Version:
+		salt = quicV2Salt
+		initialLabel = "client in"
+		keyLabel = "quicv2 key"
+		ivLabel = "quicv2 iv"
+		hpLabel = "quicv2 hp"
+	default:
+		return "", fmt.Errorf("unsupported QUIC version: 0x%08x", version)
+	}
+
+	// Parse Long Header
+	if len(data) < 6 {
+		return "", fmt.Errorf("packet too short for DCID length")
+	}
+	dcidLen := int(data[5])
+	if len(data) < 6+dcidLen+1 {
+		return "", fmt.Errorf("packet too short for DCID")
+	}
+	dcid := data[6 : 6+dcidLen]
+
+	scidLenOff := 6 + dcidLen
+	scidLen := int(data[scidLenOff])
+	tokenLenOff := scidLenOff + 1 + scidLen
+
+	if tokenLenOff >= len(data) {
+		return "", fmt.Errorf("packet too short for token length")
+	}
+
+	// Token length is a variable-length integer
+	tokenLen, tokenLenSize, err := readVarInt(data[tokenLenOff:])
+	if err != nil {
+		return "", fmt.Errorf("read token length: %w", err)
+	}
+
+	payloadLenOff := tokenLenOff + tokenLenSize + int(tokenLen)
+	if payloadLenOff >= len(data) {
+		return "", fmt.Errorf("packet too short for payload length")
+	}
+
+	// Payload length is a variable-length integer
+	payloadLen, payloadLenSize, err := readVarInt(data[payloadLenOff:])
+	if err != nil {
+		return "", fmt.Errorf("read payload length: %w", err)
+	}
+
+	pnOffset := payloadLenOff + payloadLenSize
+	if pnOffset+4 > len(data) {
+		return "", fmt.Errorf("packet too short for packet number")
+	}
+
+	// Derive initial keys
+	clientKey, clientIV, clientHP, err := deriveInitialKeys(dcid, salt, initialLabel, keyLabel, ivLabel, hpLabel)
+	if err != nil {
+		return "", fmt.Errorf("derive initial keys: %w", err)
+	}
+
+	// Remove header protection
+	sampleOffset := pnOffset + 4 // sample starts 4 bytes after pn offset
+	if sampleOffset+16 > len(data) {
+		return "", fmt.Errorf("packet too short for HP sample")
+	}
+	sample := data[sampleOffset : sampleOffset+16]
+
+	hpBlock, err := aes.NewCipher(clientHP)
+	if err != nil {
+		return "", fmt.Errorf("create HP cipher: %w", err)
+	}
+
+	mask := make([]byte, 16)
+	hpBlock.Encrypt(mask, sample)
+
+	// Unmask header byte
+	header := make([]byte, len(data))
+	copy(header, data)
+	header[0] ^= mask[0] & 0x0f // Long header: low 4 bits
+
+	// Determine packet number length
+	pnLen := int(header[0]&0x03) + 1
+
+	// Unmask packet number
+	for i := 0; i < pnLen; i++ {
+		header[pnOffset+i] ^= mask[1+i]
+	}
+
+	// Reconstruct packet number
+	var pn uint32
+	for i := 0; i < pnLen; i++ {
+		pn = (pn << 8) | uint32(header[pnOffset+i])
+	}
+
+	// Build nonce
+	nonce := make([]byte, len(clientIV))
+	copy(nonce, clientIV)
+	for i := 0; i < 4; i++ {
+		nonce[len(nonce)-1-i] ^= byte(pn >> (8 * i))
+	}
+
+	// Decrypt payload
+	block, err := aes.NewCipher(clientKey)
+	if err != nil {
+		return "", fmt.Errorf("create AES cipher: %w", err)
+	}
+
+	aead, err := cipher.NewGCM(block)
+	if err != nil {
+		return "", fmt.Errorf("create AEAD: %w", err)
+	}
+
+	encryptedPayload := header[pnOffset+pnLen : pnOffset+int(payloadLen)]
+	aad := header[:pnOffset+pnLen]
+
+	plaintext, err := aead.Open(nil, nonce, encryptedPayload, aad)
+	if err != nil {
+		return "", fmt.Errorf("decrypt QUIC payload: %w", err)
+	}
+
+	// Parse CRYPTO frames to extract ClientHello
+	clientHello, err := extractCryptoFrames(plaintext)
+	if err != nil {
+		return "", fmt.Errorf("extract CRYPTO frames: %w", err)
+	}
+
+	info, err := parseHelloBody(clientHello)
+	return info.SNI, err
+}
+
+// deriveInitialKeys derives the client's initial encryption keys from the DCID.
+func deriveInitialKeys(dcid, salt []byte, initialLabel, keyLabel, ivLabel, hpLabel string) (key, iv, hp []byte, err error) {
+	// initial_secret = HKDF-Extract(salt, DCID)
+	initialSecret := hkdf.Extract(sha256.New, dcid, salt)
+
+	// client_initial_secret = HKDF-Expand-Label(initial_secret, initialLabel, "", 32)
+	clientSecret, err := hkdfExpandLabel(initialSecret, initialLabel, nil, 32)
+	if err != nil {
+		return nil, nil, nil, fmt.Errorf("derive client secret: %w", err)
+	}
+
+	// client_key = HKDF-Expand-Label(client_secret, keyLabel, "", 16)
+	key, err = hkdfExpandLabel(clientSecret, keyLabel, nil, 16)
+	if err != nil {
+		return nil, nil, nil, fmt.Errorf("derive key: %w", err)
+	}
+
+	// client_iv = HKDF-Expand-Label(client_secret, ivLabel, "", 12)
+	iv, err = hkdfExpandLabel(clientSecret, ivLabel, nil, 12)
+	if err != nil {
+		return nil, nil, nil, fmt.Errorf("derive IV: %w", err)
+	}
+
+	// client_hp = HKDF-Expand-Label(client_secret, hpLabel, "", 16)
+	hp, err = hkdfExpandLabel(clientSecret, hpLabel, nil, 16)
+	if err != nil {
+		return nil, nil, nil, fmt.Errorf("derive HP key: %w", err)
+	}
+
+	return key, iv, hp, nil
+}
+
+// hkdfExpandLabel implements TLS 1.3 HKDF-Expand-Label.
+func hkdfExpandLabel(secret []byte, label string, context []byte, length int) ([]byte, error) {
+	// HkdfLabel = struct {
+	//   uint16 length;
+	//   opaque label<7..255> = "tls13 " + Label;
+	//   opaque context<0..255> = Context;
+	// }
+	fullLabel := "tls13 " + label
+
+	hkdfLabel := make([]byte, 2+1+len(fullLabel)+1+len(context))
+	binary.BigEndian.PutUint16(hkdfLabel[0:2], uint16(length))
+	hkdfLabel[2] = byte(len(fullLabel))
+	copy(hkdfLabel[3:], fullLabel)
+	hkdfLabel[3+len(fullLabel)] = byte(len(context))
+	if len(context) > 0 {
+		copy(hkdfLabel[4+len(fullLabel):], context)
+	}
+
+	expander := hkdf.Expand(sha256.New, secret, hkdfLabel)
+	out := make([]byte, length)
+	if _, err := io.ReadFull(expander, out); err != nil {
+		return nil, err
+	}
+	return out, nil
+}
+
+// maxCryptoFrameSize limits total CRYPTO frame data to prevent memory exhaustion.
+const maxCryptoFrameSize = 64 * 1024
+
+// extractCryptoFrames reassembles CRYPTO frame data from QUIC frames.
+func extractCryptoFrames(frames []byte) ([]byte, error) {
+	var result []byte
+	pos := 0
+
+	for pos < len(frames) {
+		frameType := frames[pos]
+
+		switch {
+		case frameType == 0x00:
+			// PADDING frame
+			pos++
+
+		case frameType == 0x06:
+			// CRYPTO frame
+			pos++
+
+			offset, n, err := readVarInt(frames[pos:])
+			if err != nil {
+				return nil, fmt.Errorf("read crypto offset: %w", err)
+			}
+			pos += n
+			_ = offset // We assume ordered, offset 0 for Initial
+
+			dataLen, n, err := readVarInt(frames[pos:])
+			if err != nil {
+				return nil, fmt.Errorf("read crypto data length: %w", err)
+			}
+			pos += n
+
+			end := pos + int(dataLen)
+			if end > len(frames) {
+				return nil, fmt.Errorf("CRYPTO frame data truncated")
+			}
+
+			result = append(result, frames[pos:end]...)
+			if len(result) > maxCryptoFrameSize {
+				return nil, fmt.Errorf("CRYPTO frame data exceeds %d bytes", maxCryptoFrameSize)
+			}
+			pos = end
+
+		case frameType == 0x01:
+			// PING frame
+			pos++
+
+		case frameType == 0x02 || frameType == 0x03:
+			// ACK frame - skip
+			pos++
+			// Largest Acknowledged
+			_, n, err := readVarInt(frames[pos:])
+			if err != nil {
+				return nil, fmt.Errorf("read ACK: %w", err)
+			}
+			pos += n
+			// ACK Delay
+			_, n, err = readVarInt(frames[pos:])
+			if err != nil {
+				return nil, fmt.Errorf("read ACK delay: %w", err)
+			}
+			pos += n
+			// ACK Range Count
+			rangeCount, n, err := readVarInt(frames[pos:])
+			if err != nil {
+				return nil, fmt.Errorf("read ACK range count: %w", err)
+			}
+			pos += n
+			// First ACK Range
+			_, n, err = readVarInt(frames[pos:])
+			if err != nil {
+				return nil, fmt.Errorf("read first ACK range: %w", err)
+			}
+			pos += n
+			// Additional ranges
+			for i := uint64(0); i < rangeCount; i++ {
+				_, n, err = readVarInt(frames[pos:])
+				if err != nil {
+					return nil, fmt.Errorf("read ACK gap: %w", err)
+				}
+				pos += n
+				_, n, err = readVarInt(frames[pos:])
+				if err != nil {
+					return nil, fmt.Errorf("read ACK range: %w", err)
+				}
+				pos += n
+			}
+			// ECN counts for type 0x03
+			if frameType == 0x03 {
+				for range 3 {
+					_, n, err = readVarInt(frames[pos:])
+					if err != nil {
+						return nil, fmt.Errorf("read ECN count: %w", err)
+					}
+					pos += n
+				}
+			}
+
+		default:
+			// Unknown frame type, stop parsing
+			if len(result) > 0 {
+				return result, nil
+			}
+			return nil, fmt.Errorf("unknown QUIC frame type: 0x%02x at offset %d", frameType, pos)
+		}
+	}
+
+	if len(result) == 0 {
+		return nil, fmt.Errorf("no CRYPTO frames found")
+	}
+
+	return result, nil
+}
+
+// readVarInt reads a QUIC variable-length integer.
+// Returns (value, bytes consumed, error).
+func readVarInt(data []byte) (uint64, int, error) {
+	if len(data) == 0 {
+		return 0, 0, fmt.Errorf("empty data for varint")
+	}
+
+	prefix := data[0] >> 6
+	length := 1 << prefix
+
+	if len(data) < length {
+		return 0, 0, fmt.Errorf("varint truncated: need %d, have %d", length, len(data))
+	}
+
+	var val uint64
+	switch length {
+	case 1:
+		val = uint64(data[0] & 0x3f)
+	case 2:
+		val = uint64(binary.BigEndian.Uint16(data[:2])) & 0x3fff
+	case 4:
+		val = uint64(binary.BigEndian.Uint32(data[:4])) & 0x3fffffff
+	case 8:
+		val = binary.BigEndian.Uint64(data[:8]) & 0x3fffffffffffffff
+	}
+
+	return val, length, nil
+}

client/inspect/quic_test.go

@@ -0,0 +1,99 @@
+package inspect
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestReadVarInt(t *testing.T) {
+	tests := []struct {
+		name string
+		data []byte
+		want uint64
+		n    int
+	}{
+		{
+			name: "1 byte value",
+			data: []byte{0x25},
+			want: 37,
+			n:    1,
+		},
+		{
+			name: "2 byte value",
+			data: []byte{0x7b, 0xbd},
+			want: 15293,
+			n:    2,
+		},
+		{
+			name: "4 byte value",
+			data: []byte{0x9d, 0x7f, 0x3e, 0x7d},
+			want: 494878333,
+			n:    4,
+		},
+		{
+			name: "zero",
+			data: []byte{0x00},
+			want: 0,
+			n:    1,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			val, n, err := readVarInt(tt.data)
+			require.NoError(t, err)
+			assert.Equal(t, tt.want, val)
+			assert.Equal(t, tt.n, n)
+		})
+	}
+}
+
+func TestReadVarInt_Empty(t *testing.T) {
+	_, _, err := readVarInt(nil)
+	require.Error(t, err)
+}
+
+func TestReadVarInt_Truncated(t *testing.T) {
+	// 2-byte prefix but only 1 byte
+	_, _, err := readVarInt([]byte{0x40})
+	require.Error(t, err)
+}
+
+func TestExtractQUICSNI_NotLongHeader(t *testing.T) {
+	// Short header packet (form bit not set)
+	data := make([]byte, 100)
+	data[0] = 0x40 // short header
+
+	_, err := ExtractQUICSNI(data)
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "not a QUIC long header")
+}
+
+func TestExtractQUICSNI_UnsupportedVersion(t *testing.T) {
+	data := make([]byte, 100)
+	data[0] = 0xC0 // long header
+	// Version 0xdeadbeef
+	data[1] = 0xde
+	data[2] = 0xad
+	data[3] = 0xbe
+	data[4] = 0xef
+
+	_, err := ExtractQUICSNI(data)
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "unsupported QUIC version")
+}
+
+func TestExtractQUICSNI_TooShort(t *testing.T) {
+	_, err := ExtractQUICSNI([]byte{0xC0, 0x00})
+	require.Error(t, err)
+}
+
+func TestHkdfExpandLabel(t *testing.T) {
+	// Smoke test: ensure it returns the right length and doesn't error
+	secret := make([]byte, 32)
+	result, err := hkdfExpandLabel(secret, "quic key", nil, 16)
+	require.NoError(t, err)
+	assert.Len(t, result, 16)
+}

client/inspect/rules.go

@@ -0,0 +1,253 @@
+package inspect
+
+import (
+	"net/netip"
+	"slices"
+	"sort"
+	"strings"
+	"sync"
+
+	log "github.com/sirupsen/logrus"
+
+	"github.com/netbirdio/netbird/client/internal/acl/id"
+	"github.com/netbirdio/netbird/shared/management/domain"
+)
+
+// RuleEngine evaluates proxy rules against connection metadata.
+// It is safe for concurrent use.
+type RuleEngine struct {
+	mu    sync.RWMutex
+	rules []Rule
+	// defaultAction applies when no rule matches.
+	defaultAction Action
+	log           *log.Entry
+}
+
+// NewRuleEngine creates a rule engine with the given default action.
+func NewRuleEngine(logger *log.Entry, defaultAction Action) *RuleEngine {
+	return &RuleEngine{
+		defaultAction: defaultAction,
+		log:           logger,
+	}
+}
+
+// UpdateRules replaces the rule set and default action. Rules are sorted by priority.
+func (e *RuleEngine) UpdateRules(rules []Rule, defaultAction Action) {
+	sorted := make([]Rule, len(rules))
+	copy(sorted, rules)
+	sort.Slice(sorted, func(i, j int) bool {
+		return sorted[i].Priority < sorted[j].Priority
+	})
+
+	e.mu.Lock()
+	e.rules = sorted
+	e.defaultAction = defaultAction
+	e.mu.Unlock()
+}
+
+// EvalResult holds the outcome of a rule evaluation.
+type EvalResult struct {
+	Action Action
+	RuleID id.RuleID
+}
+
+// Evaluate determines the action for a connection based on the rule set.
+// Pass empty path for connection-level evaluation (TLS/SNI), non-empty for request-level (HTTP).
+func (e *RuleEngine) Evaluate(src netip.Addr, dstDomain domain.Domain, dstAddr netip.Addr, dstPort uint16, proto ProtoType, path string) Action {
+	r := e.EvaluateWithResult(src, dstDomain, dstAddr, dstPort, proto, path)
+	return r.Action
+}
+
+// EvaluateWithResult is like Evaluate but also returns the matched rule ID.
+func (e *RuleEngine) EvaluateWithResult(src netip.Addr, dstDomain domain.Domain, dstAddr netip.Addr, dstPort uint16, proto ProtoType, path string) EvalResult {
+	e.mu.RLock()
+	defer e.mu.RUnlock()
+
+	for i := range e.rules {
+		rule := &e.rules[i]
+		if e.ruleMatches(rule, src, dstDomain, dstAddr, dstPort, proto, path) {
+			e.log.Tracef("rule %s matched: action=%s src=%s domain=%s dst=%s:%d proto=%s path=%s",
+				rule.ID, rule.Action, src, dstDomain.SafeString(), dstAddr, dstPort, proto, path)
+			return EvalResult{Action: rule.Action, RuleID: rule.ID}
+		}
+	}
+
+	e.log.Tracef("no rule matched, default=%s: src=%s domain=%s dst=%s:%d proto=%s path=%s",
+		e.defaultAction, src, dstDomain.SafeString(), dstAddr, dstPort, proto, path)
+	return EvalResult{Action: e.defaultAction}
+}
+
+// HasPathRulesForDomain returns true if any rule matching the domain has non-empty Paths.
+// Used to force MITM inspection when path-level rules exist (paths are only visible after decryption).
+func (e *RuleEngine) HasPathRulesForDomain(dstDomain domain.Domain) bool {
+	e.mu.RLock()
+	defer e.mu.RUnlock()
+
+	for i := range e.rules {
+		if len(e.rules[i].Paths) > 0 && e.matchDomain(&e.rules[i], dstDomain) {
+			return true
+		}
+	}
+	return false
+}
+
+// ruleMatches checks whether all non-empty fields of a rule match.
+// Empty fields are treated as "match any".
+// All specified fields must match (AND logic).
+func (e *RuleEngine) ruleMatches(rule *Rule, src netip.Addr, dstDomain domain.Domain, dstAddr netip.Addr, dstPort uint16, proto ProtoType, path string) bool {
+	if !e.matchSource(rule, src) {
+		return false
+	}
+
+	if !e.matchDomain(rule, dstDomain) {
+		return false
+	}
+
+	if !e.matchNetwork(rule, dstAddr) {
+		return false
+	}
+
+	if !e.matchPort(rule, dstPort) {
+		return false
+	}
+
+	if !e.matchProtocol(rule, proto) {
+		return false
+	}
+
+	if !e.matchPaths(rule, path) {
+		return false
+	}
+
+	return true
+}
+
+// matchSource returns true if src matches any of the rule's source CIDRs,
+// or if no source CIDRs are specified (match any).
+func (e *RuleEngine) matchSource(rule *Rule, src netip.Addr) bool {
+	if len(rule.Sources) == 0 {
+		return true
+	}
+
+	for _, prefix := range rule.Sources {
+		if prefix.Contains(src) {
+			return true
+		}
+	}
+
+	return false
+}
+
+// matchDomain returns true if dstDomain matches any of the rule's domain patterns,
+// or if no domain patterns are specified (match any).
+func (e *RuleEngine) matchDomain(rule *Rule, dstDomain domain.Domain) bool {
+	if len(rule.Domains) == 0 {
+		return true
+	}
+
+	// If we have domain rules but no domain to match against (e.g., raw IP connection),
+	// the domain condition does not match.
+	if dstDomain == "" {
+		return false
+	}
+
+	for _, pattern := range rule.Domains {
+		if MatchDomain(pattern, dstDomain) {
+			return true
+		}
+	}
+
+	return false
+}
+
+// matchNetwork returns true if dstAddr is within any of the rule's destination CIDRs,
+// or if no destination CIDRs are specified (match any).
+func (e *RuleEngine) matchNetwork(rule *Rule, dstAddr netip.Addr) bool {
+	if len(rule.Networks) == 0 {
+		return true
+	}
+
+	for _, prefix := range rule.Networks {
+		if prefix.Contains(dstAddr) {
+			return true
+		}
+	}
+
+	return false
+}
+
+// matchProtocol returns true if proto matches any of the rule's protocols,
+// or if no protocols are specified (match any).
+func (e *RuleEngine) matchProtocol(rule *Rule, proto ProtoType) bool {
+	if len(rule.Protocols) == 0 {
+		return true
+	}
+
+	for _, p := range rule.Protocols {
+		if p == proto {
+			return true
+		}
+	}
+
+	return false
+}
+
+// matchPort returns true if dstPort matches any of the rule's destination ports,
+// or if no ports are specified (match any).
+func (e *RuleEngine) matchPort(rule *Rule, dstPort uint16) bool {
+	if len(rule.Ports) == 0 {
+		return true
+	}
+
+	return slices.Contains(rule.Ports, dstPort)
+}
+
+// matchPaths returns true if path matches any of the rule's path patterns,
+// or if no paths are specified (match any). Empty path (connection-level eval) matches all.
+func (e *RuleEngine) matchPaths(rule *Rule, path string) bool {
+	if len(rule.Paths) == 0 {
+		return true
+	}
+	// Connection-level (path=""): rules with paths don't match at connection level.
+	// HasPathRulesForDomain forces the connection to inspect, so paths are
+	// checked per-request once the HTTP request is visible.
+	if path == "" {
+		return false
+	}
+	for _, pattern := range rule.Paths {
+		if matchPath(pattern, path) {
+			return true
+		}
+	}
+	return false
+}
+
+// matchPath checks if a URL path matches a pattern.
+// Supports: exact ("/login"), prefix with wildcard ("/api/*"),
+// and contains ("*/admin/*"). A bare "*" matches everything.
+func matchPath(pattern, path string) bool {
+	if pattern == "*" {
+		return true
+	}
+
+	hasLeadingStar := strings.HasPrefix(pattern, "*")
+	hasTrailingStar := strings.HasSuffix(pattern, "*")
+
+	switch {
+	case hasLeadingStar && hasTrailingStar:
+		// */admin/* = contains
+		middle := strings.Trim(pattern, "*")
+		return strings.Contains(path, middle)
+	case hasTrailingStar:
+		// /api/* = prefix
+		prefix := strings.TrimSuffix(pattern, "*")
+		return strings.HasPrefix(path, prefix)
+	case hasLeadingStar:
+		// *.json = suffix
+		suffix := strings.TrimPrefix(pattern, "*")
+		return strings.HasSuffix(path, suffix)
+	default:
+		// exact
+		return path == pattern
+	}
+}

client/inspect/rules_test.go

@@ -0,0 +1,338 @@
+package inspect
+
+import (
+	"net/netip"
+	"testing"
+
+	log "github.com/sirupsen/logrus"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/client/internal/acl/id"
+	"github.com/netbirdio/netbird/shared/management/domain"
+)
+
+func testLogger() *log.Entry {
+	return log.WithField("test", true)
+}
+
+func mustDomain(t *testing.T, s string) domain.Domain {
+	t.Helper()
+	d, err := domain.FromString(s)
+	require.NoError(t, err)
+	return d
+}
+
+func TestRuleEngine_Evaluate(t *testing.T) {
+	tests := []struct {
+		name          string
+		rules         []Rule
+		defaultAction Action
+		src           netip.Addr
+		dstDomain     domain.Domain
+		dstAddr       netip.Addr
+		dstPort       uint16
+		want          Action
+	}{
+		{
+			name:          "no rules returns default allow",
+			defaultAction: ActionAllow,
+			src:           netip.MustParseAddr("10.0.0.1"),
+			dstAddr:       netip.MustParseAddr("1.2.3.4"),
+			dstPort:       443,
+			want:          ActionAllow,
+		},
+		{
+			name:          "no rules returns default block",
+			defaultAction: ActionBlock,
+			src:           netip.MustParseAddr("10.0.0.1"),
+			dstAddr:       netip.MustParseAddr("1.2.3.4"),
+			dstPort:       443,
+			want:          ActionBlock,
+		},
+		{
+			name:          "domain exact match blocks",
+			defaultAction: ActionAllow,
+			rules: []Rule{
+				{
+					ID:      id.RuleID("r1"),
+					Domains: []domain.Domain{mustDomain(t, "malware.example.com")},
+					Action:  ActionBlock,
+				},
+			},
+			src:       netip.MustParseAddr("10.0.0.1"),
+			dstDomain: mustDomain(t, "malware.example.com"),
+			dstAddr:   netip.MustParseAddr("1.2.3.4"),
+			dstPort:   443,
+			want:      ActionBlock,
+		},
+		{
+			name:          "domain wildcard match blocks",
+			defaultAction: ActionAllow,
+			rules: []Rule{
+				{
+					ID:      id.RuleID("r1"),
+					Domains: []domain.Domain{mustDomain(t, "*.evil.com")},
+					Action:  ActionBlock,
+				},
+			},
+			src:       netip.MustParseAddr("10.0.0.1"),
+			dstDomain: mustDomain(t, "phishing.evil.com"),
+			dstAddr:   netip.MustParseAddr("1.2.3.4"),
+			dstPort:   443,
+			want:      ActionBlock,
+		},
+		{
+			name:          "domain wildcard does not match base",
+			defaultAction: ActionAllow,
+			rules: []Rule{
+				{
+					ID:      id.RuleID("r1"),
+					Domains: []domain.Domain{mustDomain(t, "*.evil.com")},
+					Action:  ActionBlock,
+				},
+			},
+			src:       netip.MustParseAddr("10.0.0.1"),
+			dstDomain: mustDomain(t, "evil.com"),
+			dstAddr:   netip.MustParseAddr("1.2.3.4"),
+			dstPort:   443,
+			want:      ActionAllow,
+		},
+		{
+			name:          "case insensitive domain match",
+			defaultAction: ActionAllow,
+			rules: []Rule{
+				{
+					ID:      id.RuleID("r1"),
+					Domains: []domain.Domain{mustDomain(t, "Example.COM")},
+					Action:  ActionBlock,
+				},
+			},
+			src:       netip.MustParseAddr("10.0.0.1"),
+			dstDomain: mustDomain(t, "EXAMPLE.com"),
+			dstAddr:   netip.MustParseAddr("1.2.3.4"),
+			dstPort:   443,
+			want:      ActionBlock,
+		},
+		{
+			name:          "source CIDR match",
+			defaultAction: ActionAllow,
+			rules: []Rule{
+				{
+					ID:      id.RuleID("r1"),
+					Sources: []netip.Prefix{netip.MustParsePrefix("192.168.1.0/24")},
+					Action:  ActionInspect,
+				},
+			},
+			src:     netip.MustParseAddr("192.168.1.50"),
+			dstAddr: netip.MustParseAddr("1.2.3.4"),
+			dstPort: 443,
+			want:    ActionInspect,
+		},
+		{
+			name:          "source CIDR no match",
+			defaultAction: ActionAllow,
+			rules: []Rule{
+				{
+					ID:      id.RuleID("r1"),
+					Sources: []netip.Prefix{netip.MustParsePrefix("192.168.1.0/24")},
+					Action:  ActionBlock,
+				},
+			},
+			src:     netip.MustParseAddr("10.0.0.5"),
+			dstAddr: netip.MustParseAddr("1.2.3.4"),
+			dstPort: 443,
+			want:    ActionAllow,
+		},
+		{
+			name:          "destination network match",
+			defaultAction: ActionAllow,
+			rules: []Rule{
+				{
+					ID:       id.RuleID("r1"),
+					Networks: []netip.Prefix{netip.MustParsePrefix("10.0.0.0/8")},
+					Action:   ActionInspect,
+				},
+			},
+			src:     netip.MustParseAddr("192.168.1.1"),
+			dstAddr: netip.MustParseAddr("10.50.0.1"),
+			dstPort: 80,
+			want:    ActionInspect,
+		},
+		{
+			name:          "port match",
+			defaultAction: ActionAllow,
+			rules: []Rule{
+				{
+					ID:     id.RuleID("r1"),
+					Ports:  []uint16{443, 8443},
+					Action: ActionInspect,
+				},
+			},
+			src:     netip.MustParseAddr("10.0.0.1"),
+			dstAddr: netip.MustParseAddr("1.2.3.4"),
+			dstPort: 443,
+			want:    ActionInspect,
+		},
+		{
+			name:          "port no match",
+			defaultAction: ActionAllow,
+			rules: []Rule{
+				{
+					ID:     id.RuleID("r1"),
+					Ports:  []uint16{443, 8443},
+					Action: ActionBlock,
+				},
+			},
+			src:     netip.MustParseAddr("10.0.0.1"),
+			dstAddr: netip.MustParseAddr("1.2.3.4"),
+			dstPort: 22,
+			want:    ActionAllow,
+		},
+		{
+			name:          "priority ordering first match wins",
+			defaultAction: ActionAllow,
+			rules: []Rule{
+				{
+					ID:       id.RuleID("allow-internal"),
+					Domains:  []domain.Domain{mustDomain(t, "*.internal.corp")},
+					Action:   ActionAllow,
+					Priority: 1,
+				},
+				{
+					ID:       id.RuleID("inspect-all"),
+					Action:   ActionInspect,
+					Priority: 10,
+				},
+			},
+			src:       netip.MustParseAddr("10.0.0.1"),
+			dstDomain: mustDomain(t, "api.internal.corp"),
+			dstAddr:   netip.MustParseAddr("10.1.0.5"),
+			dstPort:   443,
+			want:      ActionAllow,
+		},
+		{
+			name:          "all fields must match (AND logic)",
+			defaultAction: ActionAllow,
+			rules: []Rule{
+				{
+					ID:      id.RuleID("r1"),
+					Sources: []netip.Prefix{netip.MustParsePrefix("192.168.1.0/24")},
+					Domains: []domain.Domain{mustDomain(t, "*.evil.com")},
+					Ports:   []uint16{443},
+					Action:  ActionBlock,
+				},
+			},
+			// Source matches, domain matches, but port doesn't
+			src:       netip.MustParseAddr("192.168.1.10"),
+			dstDomain: mustDomain(t, "phish.evil.com"),
+			dstAddr:   netip.MustParseAddr("1.2.3.4"),
+			dstPort:   8080,
+			want:      ActionAllow,
+		},
+		{
+			name:          "empty domain with domain rule does not match",
+			defaultAction: ActionAllow,
+			rules: []Rule{
+				{
+					ID:      id.RuleID("r1"),
+					Domains: []domain.Domain{mustDomain(t, "example.com")},
+					Action:  ActionBlock,
+				},
+			},
+			src:       netip.MustParseAddr("10.0.0.1"),
+			dstDomain: "", // raw IP connection, no SNI
+			dstAddr:   netip.MustParseAddr("1.2.3.4"),
+			dstPort:   443,
+			want:      ActionAllow,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			engine := NewRuleEngine(testLogger(), tt.defaultAction)
+			engine.UpdateRules(tt.rules, tt.defaultAction)
+
+			got := engine.Evaluate(tt.src, tt.dstDomain, tt.dstAddr, tt.dstPort, "", "")
+			assert.Equal(t, tt.want, got)
+		})
+	}
+}
+
+func TestRuleEngine_ProtocolMatching(t *testing.T) {
+	engine := NewRuleEngine(testLogger(), ActionAllow)
+	engine.UpdateRules([]Rule{
+		{
+			ID:        "block-websocket",
+			Protocols: []ProtoType{ProtoWebSocket},
+			Action:    ActionBlock,
+			Priority:  1,
+		},
+		{
+			ID:        "inspect-h2",
+			Protocols: []ProtoType{ProtoH2},
+			Action:    ActionInspect,
+			Priority:  2,
+		},
+	}, ActionAllow)
+
+	src := netip.MustParseAddr("10.0.0.1")
+	dst := netip.MustParseAddr("1.2.3.4")
+
+	// WebSocket: blocked by rule
+	assert.Equal(t, ActionBlock, engine.Evaluate(src, "", dst, 443, ProtoWebSocket, ""))
+
+	// HTTP/2: inspected by rule
+	assert.Equal(t, ActionInspect, engine.Evaluate(src, "", dst, 443, ProtoH2, ""))
+
+	// Plain HTTP: no protocol rule matches, default allow
+	assert.Equal(t, ActionAllow, engine.Evaluate(src, "", dst, 80, ProtoHTTP, ""))
+
+	// HTTPS: no protocol rule matches, default allow
+	assert.Equal(t, ActionAllow, engine.Evaluate(src, "", dst, 443, ProtoHTTPS, ""))
+
+	// QUIC/H3: no protocol rule matches, default allow
+	assert.Equal(t, ActionAllow, engine.Evaluate(src, "", dst, 443, ProtoH3, ""))
+
+	// Empty protocol (unknown): no protocol rule matches, default allow
+	assert.Equal(t, ActionAllow, engine.Evaluate(src, "", dst, 443, "", ""))
+}
+
+func TestRuleEngine_EmptyProtocolsMatchAll(t *testing.T) {
+	engine := NewRuleEngine(testLogger(), ActionAllow)
+	engine.UpdateRules([]Rule{
+		{
+			ID:     "block-all-protos",
+			Action: ActionBlock,
+			// No Protocols field = match all protocols
+			Priority: 1,
+		},
+	}, ActionAllow)
+
+	src := netip.MustParseAddr("10.0.0.1")
+	dst := netip.MustParseAddr("1.2.3.4")
+
+	assert.Equal(t, ActionBlock, engine.Evaluate(src, "", dst, 443, ProtoHTTP, ""))
+	assert.Equal(t, ActionBlock, engine.Evaluate(src, "", dst, 443, ProtoHTTPS, ""))
+	assert.Equal(t, ActionBlock, engine.Evaluate(src, "", dst, 443, ProtoWebSocket, ""))
+	assert.Equal(t, ActionBlock, engine.Evaluate(src, "", dst, 443, ProtoH2, ""))
+	assert.Equal(t, ActionBlock, engine.Evaluate(src, "", dst, 443, "", ""))
+}
+
+func TestRuleEngine_UpdateRulesSortsByPriority(t *testing.T) {
+	engine := NewRuleEngine(testLogger(), ActionAllow)
+
+	engine.UpdateRules([]Rule{
+		{ID: "c", Priority: 30, Action: ActionBlock},
+		{ID: "a", Priority: 10, Action: ActionInspect},
+		{ID: "b", Priority: 20, Action: ActionAllow},
+	}, ActionAllow)
+
+	engine.mu.RLock()
+	defer engine.mu.RUnlock()
+
+	require.Len(t, engine.rules, 3)
+	assert.Equal(t, id.RuleID("a"), engine.rules[0].ID)
+	assert.Equal(t, id.RuleID("b"), engine.rules[1].ID)
+	assert.Equal(t, id.RuleID("c"), engine.rules[2].ID)
+}

client/inspect/sni.go

@@ -0,0 +1,287 @@
+package inspect
+
+import (
+	"encoding/binary"
+	"fmt"
+	"io"
+
+	"github.com/netbirdio/netbird/shared/management/domain"
+)
+
+const (
+	recordTypeHandshake      = 0x16
+	handshakeTypeClientHello = 0x01
+	extensionTypeSNI         = 0x0000
+	extensionTypeALPN        = 0x0010
+	sniTypeHostName          = 0x00
+
+	// maxClientHelloSize is the maximum ClientHello size we'll read.
+	// Real-world ClientHellos are typically under 1KB but can reach ~16KB with
+	// many extensions (post-quantum key shares, etc.).
+	maxClientHelloSize = 16384
+)
+
+// ClientHelloInfo holds data extracted from a TLS ClientHello.
+type ClientHelloInfo struct {
+	SNI  domain.Domain
+	ALPN []string
+}
+
+// isTLSHandshake reports whether the first byte indicates a TLS handshake record.
+func isTLSHandshake(b byte) bool {
+	return b == recordTypeHandshake
+}
+
+// httpMethods lists the first bytes of valid HTTP method tokens.
+var httpMethods = [][]byte{
+	[]byte("GET "),
+	[]byte("POST"),
+	[]byte("PUT "),
+	[]byte("DELE"),
+	[]byte("HEAD"),
+	[]byte("OPTI"),
+	[]byte("PATC"),
+	[]byte("CONN"),
+	[]byte("TRAC"),
+}
+
+// isHTTPMethod reports whether the peeked bytes look like the start of an HTTP request.
+func isHTTPMethod(b []byte) bool {
+	if len(b) < 4 {
+		return false
+	}
+	for _, m := range httpMethods {
+		if b[0] == m[0] && b[1] == m[1] && b[2] == m[2] && b[3] == m[3] {
+			return true
+		}
+	}
+	return false
+}
+
+// parseClientHello reads a TLS ClientHello from r and returns SNI and ALPN.
+func parseClientHello(r io.Reader) (ClientHelloInfo, error) {
+	// TLS record header: type(1) + version(2) + length(2)
+	var recordHeader [5]byte
+	if _, err := io.ReadFull(r, recordHeader[:]); err != nil {
+		return ClientHelloInfo{}, fmt.Errorf("read TLS record header: %w", err)
+	}
+
+	if recordHeader[0] != recordTypeHandshake {
+		return ClientHelloInfo{}, fmt.Errorf("not a TLS handshake record (type=%d)", recordHeader[0])
+	}
+
+	recordLen := int(binary.BigEndian.Uint16(recordHeader[3:5]))
+	if recordLen < 4 || recordLen > maxClientHelloSize {
+		return ClientHelloInfo{}, fmt.Errorf("invalid TLS record length: %d", recordLen)
+	}
+
+	// Read the full handshake message
+	msg := make([]byte, recordLen)
+	if _, err := io.ReadFull(r, msg); err != nil {
+		return ClientHelloInfo{}, fmt.Errorf("read handshake message: %w", err)
+	}
+
+	return parseClientHelloMsg(msg)
+}
+
+// extractSNI reads a TLS ClientHello from r and returns the SNI hostname.
+// Returns empty domain if no SNI extension is present.
+func extractSNI(r io.Reader) (domain.Domain, error) {
+	info, err := parseClientHello(r)
+	return info.SNI, err
+}
+
+// extractSNIFromBytes parses SNI from raw bytes that start with the TLS record header.
+func extractSNIFromBytes(data []byte) (domain.Domain, error) {
+	info, err := parseClientHelloFromBytes(data)
+	return info.SNI, err
+}
+
+// parseClientHelloFromBytes parses a ClientHello from raw bytes starting with the TLS record header.
+func parseClientHelloFromBytes(data []byte) (ClientHelloInfo, error) {
+	if len(data) < 5 {
+		return ClientHelloInfo{}, fmt.Errorf("data too short for TLS record header")
+	}
+
+	if data[0] != recordTypeHandshake {
+		return ClientHelloInfo{}, fmt.Errorf("not a TLS handshake record (type=%d)", data[0])
+	}
+
+	recordLen := int(binary.BigEndian.Uint16(data[3:5]))
+	if recordLen < 4 {
+		return ClientHelloInfo{}, fmt.Errorf("invalid TLS record length: %d", recordLen)
+	}
+
+	end := 5 + recordLen
+	if end > len(data) {
+		return ClientHelloInfo{}, fmt.Errorf("TLS record truncated: need %d, have %d", end, len(data))
+	}
+
+	return parseClientHelloMsg(data[5:end])
+}
+
+// parseClientHelloMsg extracts SNI and ALPN from a raw ClientHello handshake message.
+// msg starts at the handshake type byte.
+func parseClientHelloMsg(msg []byte) (ClientHelloInfo, error) {
+	if len(msg) < 4 {
+		return ClientHelloInfo{}, fmt.Errorf("handshake message too short")
+	}
+
+	if msg[0] != handshakeTypeClientHello {
+		return ClientHelloInfo{}, fmt.Errorf("not a ClientHello (type=%d)", msg[0])
+	}
+
+	// Handshake header: type(1) + length(3)
+	helloLen := int(msg[1])<<16 | int(msg[2])<<8 | int(msg[3])
+	if helloLen+4 > len(msg) {
+		return ClientHelloInfo{}, fmt.Errorf("ClientHello truncated")
+	}
+
+	hello := msg[4 : 4+helloLen]
+	return parseHelloBody(hello)
+}
+
+// parseHelloBody parses the ClientHello body (after handshake header)
+// and extracts SNI and ALPN.
+func parseHelloBody(hello []byte) (ClientHelloInfo, error) {
+	// ClientHello structure:
+	// version(2) + random(32) + session_id_len(1) + session_id(var)
+	// + cipher_suites_len(2) + cipher_suites(var)
+	// + compression_len(1) + compression(var)
+	// + extensions_len(2) + extensions(var)
+
+	var info ClientHelloInfo
+
+	if len(hello) < 35 {
+		return info, fmt.Errorf("ClientHello body too short")
+	}
+
+	pos := 2 + 32 // skip version + random
+
+	// Skip session ID
+	if pos >= len(hello) {
+		return info, fmt.Errorf("ClientHello truncated at session ID")
+	}
+	sessionIDLen := int(hello[pos])
+	pos += 1 + sessionIDLen
+
+	// Skip cipher suites
+	if pos+2 > len(hello) {
+		return info, fmt.Errorf("ClientHello truncated at cipher suites")
+	}
+	cipherLen := int(binary.BigEndian.Uint16(hello[pos : pos+2]))
+	pos += 2 + cipherLen
+
+	// Skip compression methods
+	if pos >= len(hello) {
+		return info, fmt.Errorf("ClientHello truncated at compression")
+	}
+	compLen := int(hello[pos])
+	pos += 1 + compLen
+
+	// Extensions
+	if pos+2 > len(hello) {
+		return info, nil
+	}
+
+	extLen := int(binary.BigEndian.Uint16(hello[pos : pos+2]))
+	pos += 2
+
+	extEnd := pos + extLen
+	if extEnd > len(hello) {
+		return info, fmt.Errorf("extensions block truncated")
+	}
+
+	// Walk extensions looking for SNI and ALPN
+	for pos+4 <= extEnd {
+		extType := binary.BigEndian.Uint16(hello[pos : pos+2])
+		extDataLen := int(binary.BigEndian.Uint16(hello[pos+2 : pos+4]))
+		pos += 4
+
+		if pos+extDataLen > extEnd {
+			return info, fmt.Errorf("extension data truncated")
+		}
+
+		switch extType {
+		case extensionTypeSNI:
+			sni, err := parseSNIExtension(hello[pos : pos+extDataLen])
+			if err != nil {
+				return info, err
+			}
+			info.SNI = sni
+		case extensionTypeALPN:
+			info.ALPN = parseALPNExtension(hello[pos : pos+extDataLen])
+		}
+
+		pos += extDataLen
+	}
+
+	return info, nil
+}
+
+// parseALPNExtension parses the ALPN extension data and returns protocol names.
+// ALPN extension: list_length(2) + entries (each: len(1) + protocol_name(var))
+func parseALPNExtension(data []byte) []string {
+	if len(data) < 2 {
+		return nil
+	}
+
+	listLen := int(binary.BigEndian.Uint16(data[0:2]))
+	if listLen+2 > len(data) {
+		return nil
+	}
+
+	var protocols []string
+	pos := 2
+	end := 2 + listLen
+
+	for pos < end {
+		if pos >= len(data) {
+			break
+		}
+		nameLen := int(data[pos])
+		pos++
+		if pos+nameLen > end {
+			break
+		}
+		protocols = append(protocols, string(data[pos:pos+nameLen]))
+		pos += nameLen
+	}
+
+	return protocols
+}
+
+// parseSNIExtension parses the SNI extension data and returns the hostname.
+func parseSNIExtension(data []byte) (domain.Domain, error) {
+	// SNI extension: list_length(2) + entries
+	if len(data) < 2 {
+		return "", fmt.Errorf("SNI extension too short")
+	}
+
+	listLen := int(binary.BigEndian.Uint16(data[0:2]))
+	if listLen+2 > len(data) {
+		return "", fmt.Errorf("SNI list truncated")
+	}
+
+	pos := 2
+	end := 2 + listLen
+
+	for pos+3 <= end {
+		nameType := data[pos]
+		nameLen := int(binary.BigEndian.Uint16(data[pos+1 : pos+3]))
+		pos += 3
+
+		if pos+nameLen > end {
+			return "", fmt.Errorf("SNI name truncated")
+		}
+
+		if nameType == sniTypeHostName {
+			hostname := string(data[pos : pos+nameLen])
+			return domain.FromString(hostname)
+		}
+
+		pos += nameLen
+	}
+
+	return "", nil
+}

client/inspect/sni_test.go

@@ -0,0 +1,109 @@
+package inspect
+
+import (
+	"bytes"
+	"crypto/tls"
+	"net"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestExtractSNI(t *testing.T) {
+	tests := []struct {
+		name    string
+		sni     string
+		wantSNI string
+		wantErr bool
+	}{
+		{
+			name:    "standard domain",
+			sni:     "example.com",
+			wantSNI: "example.com",
+		},
+		{
+			name:    "subdomain",
+			sni:     "api.staging.example.com",
+			wantSNI: "api.staging.example.com",
+		},
+		{
+			name:    "mixed case normalized to lowercase",
+			sni:     "Example.COM",
+			wantSNI: "example.com",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			clientHello := buildClientHello(t, tt.sni)
+
+			sni, err := extractSNI(bytes.NewReader(clientHello))
+			if tt.wantErr {
+				require.Error(t, err)
+				return
+			}
+
+			require.NoError(t, err)
+			assert.Equal(t, tt.wantSNI, sni.PunycodeString())
+		})
+	}
+}
+
+func TestExtractSNI_NotTLS(t *testing.T) {
+	// HTTP request instead of TLS
+	data := []byte("GET / HTTP/1.1\r\nHost: example.com\r\n\r\n")
+	_, err := extractSNI(bytes.NewReader(data))
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "not a TLS handshake")
+}
+
+func TestExtractSNI_Truncated(t *testing.T) {
+	// Just the record header, no body
+	data := []byte{0x16, 0x03, 0x01, 0x00, 0x05}
+	_, err := extractSNI(bytes.NewReader(data))
+	require.Error(t, err)
+}
+
+func TestExtractSNIFromBytes(t *testing.T) {
+	clientHello := buildClientHello(t, "test.example.com")
+
+	sni, err := extractSNIFromBytes(clientHello)
+	require.NoError(t, err)
+	assert.Equal(t, "test.example.com", sni.PunycodeString())
+}
+
+// buildClientHello generates a real TLS ClientHello with the given SNI.
+func buildClientHello(t *testing.T, serverName string) []byte {
+	t.Helper()
+
+	// Use a pipe to capture the ClientHello bytes
+	clientConn, serverConn := net.Pipe()
+
+	done := make(chan []byte, 1)
+	go func() {
+		buf := make([]byte, 4096)
+		n, _ := serverConn.Read(buf)
+		done <- buf[:n]
+		serverConn.Close()
+	}()
+
+	tlsConn := tls.Client(clientConn, &tls.Config{
+		ServerName:         serverName,
+		InsecureSkipVerify: true,
+	})
+
+	// Trigger the handshake (will fail since server isn't TLS, but we capture the ClientHello)
+	go func() {
+		_ = tlsConn.Handshake()
+		tlsConn.Close()
+	}()
+
+	clientHello := <-done
+	clientConn.Close()
+
+	require.True(t, len(clientHello) > 5, "ClientHello too short")
+	require.Equal(t, byte(0x16), clientHello[0], "not a TLS handshake record")
+
+	return clientHello
+}

client/inspect/tls.go

@@ -0,0 +1,287 @@
+package inspect
+
+import (
+	"context"
+	"crypto/tls"
+	"fmt"
+	"io"
+	"net"
+	"net/netip"
+
+	"github.com/netbirdio/netbird/shared/management/domain"
+)
+
+// handleTLS processes a TLS connection for the kernel-mode path: extracts SNI,
+// evaluates rules, and handles the connection internally.
+// In envoy mode, allowed connections are forwarded to envoy instead of direct relay.
+func (p *Proxy) handleTLS(ctx context.Context, pconn *peekConn, dst netip.AddrPort, src SourceInfo) error {
+	result, err := p.inspectTLS(ctx, pconn, dst, src)
+	if err != nil {
+		return err
+	}
+
+	if result.PassthroughConn != nil {
+		p.mu.RLock()
+		envoy := p.envoy
+		p.mu.RUnlock()
+
+		if envoy != nil {
+			return p.forwardToEnvoy(ctx, pconn, dst, src, envoy)
+		}
+		return p.tlsPassthrough(ctx, pconn, dst, "")
+	}
+
+	return nil
+}
+
+// inspectTLS extracts SNI, evaluates rules, and returns the result.
+// For ActionAllow: returns the peekConn as PassthroughConn (caller relays).
+// For ActionBlock/ActionInspect: handles internally and returns nil PassthroughConn.
+func (p *Proxy) inspectTLS(ctx context.Context, pconn *peekConn, dst netip.AddrPort, src SourceInfo) (InspectResult, error) {
+	// The first 5 bytes (TLS record header) are already peeked.
+	// Extend to read the full TLS record so bytes remain in the buffer for passthrough.
+	peeked := pconn.Peeked()
+	recordLen := int(peeked[3])<<8 | int(peeked[4])
+	if _, err := pconn.PeekMore(5 + recordLen); err != nil {
+		return InspectResult{}, fmt.Errorf("read TLS record: %w", err)
+	}
+
+	hello, err := parseClientHelloFromBytes(pconn.Peeked())
+	if err != nil {
+		return InspectResult{}, fmt.Errorf("parse ClientHello: %w", err)
+	}
+
+	sni := hello.SNI
+	proto := protoFromALPN(hello.ALPN)
+	// Connection-level evaluation: pass empty path.
+	action := p.evaluateAction(src.IP, sni, dst, proto, "")
+
+	// If any rule for this domain has path patterns, force inspect so paths can
+	// be checked per-request after MITM decryption.
+	if action == ActionAllow && p.rules.HasPathRulesForDomain(sni) {
+		p.log.Debugf("upgrading to inspect for %s (path rules exist)", sni.PunycodeString())
+		action = ActionInspect
+	}
+
+	// Snapshot cert provider under lock for use in this connection.
+	p.mu.RLock()
+	certs := p.certs
+	p.mu.RUnlock()
+
+	switch action {
+	case ActionBlock:
+		p.log.Debugf("block: TLS to %s (SNI=%s)", dst, sni.PunycodeString())
+		if certs != nil {
+			return InspectResult{Action: ActionBlock}, p.tlsBlockPage(ctx, pconn, sni, certs)
+		}
+		return InspectResult{Action: ActionBlock}, ErrBlocked
+
+	case ActionAllow:
+		p.log.Tracef("allow: TLS passthrough to %s (SNI=%s)", dst, sni.PunycodeString())
+		return InspectResult{Action: ActionAllow, PassthroughConn: pconn}, nil
+
+	case ActionInspect:
+		if certs == nil {
+			p.log.Warnf("allow: %s (inspect requested but no MITM CA configured)", sni.PunycodeString())
+			return InspectResult{Action: ActionAllow, PassthroughConn: pconn}, nil
+		}
+		err := p.tlsMITM(ctx, pconn, dst, sni, src, certs)
+		return InspectResult{Action: ActionInspect}, err
+
+	default:
+		p.log.Warnf("block: unknown action %q for %s", action, sni.PunycodeString())
+		return InspectResult{Action: ActionBlock}, ErrBlocked
+	}
+}
+
+// tlsBlockPage completes a MITM TLS handshake with the client using a dynamic
+// certificate, then serves an HTTP 403 block page so the user sees a clear
+// message instead of a cryptic SSL error.
+func (p *Proxy) tlsBlockPage(ctx context.Context, pconn *peekConn, sni domain.Domain, certs *CertProvider) error {
+	hostname := sni.PunycodeString()
+
+	// Force HTTP/1.1 only: block pages are simple responses, no need for h2
+	tlsCfg := certs.GetTLSConfig()
+	tlsCfg.NextProtos = []string{"http/1.1"}
+	clientTLS := tls.Server(pconn, tlsCfg)
+	if err := clientTLS.HandshakeContext(ctx); err != nil {
+		// Client may not trust our CA, handshake fails. That's expected.
+		return fmt.Errorf("block page TLS handshake for %s: %w", hostname, err)
+	}
+	defer func() {
+		if err := clientTLS.Close(); err != nil {
+			p.log.Debugf("close block page TLS for %s: %v", hostname, err)
+		}
+	}()
+
+	writeBlockResponse(clientTLS, nil, sni)
+	return ErrBlocked
+}
+
+// tlsPassthrough connects to the destination and relays encrypted traffic
+// without decryption. The peeked ClientHello bytes are replayed.
+func (p *Proxy) tlsPassthrough(ctx context.Context, pconn *peekConn, dst netip.AddrPort, sni domain.Domain) error {
+	remote, err := p.dialTCP(ctx, dst)
+	if err != nil {
+		return fmt.Errorf("dial %s: %w", dst, err)
+	}
+	defer func() {
+		if err := remote.Close(); err != nil {
+			p.log.Debugf("close remote for %s: %v", dst, err)
+		}
+	}()
+
+	p.log.Tracef("allow: TLS passthrough to %s (SNI=%s)", dst, sni.PunycodeString())
+
+	return relay(ctx, pconn, remote)
+}
+
+// tlsMITM terminates the client TLS connection with a dynamic certificate,
+// establishes a new TLS connection to the real destination, and runs the
+// HTTP inspection pipeline on the decrypted traffic.
+func (p *Proxy) tlsMITM(ctx context.Context, pconn *peekConn, dst netip.AddrPort, sni domain.Domain, src SourceInfo, certs *CertProvider) error {
+	hostname := sni.PunycodeString()
+
+	// TLS handshake with client using dynamic cert
+	clientTLS := tls.Server(pconn, certs.GetTLSConfig())
+	if err := clientTLS.HandshakeContext(ctx); err != nil {
+		return fmt.Errorf("client TLS handshake for %s: %w", hostname, err)
+	}
+	defer func() {
+		if err := clientTLS.Close(); err != nil {
+			p.log.Debugf("close client TLS for %s: %v", hostname, err)
+		}
+	}()
+
+	// TLS connection to real destination
+	remoteTLS, err := p.dialTLS(ctx, dst, hostname)
+	if err != nil {
+		return fmt.Errorf("dial TLS %s (%s): %w", dst, hostname, err)
+	}
+	defer func() {
+		if err := remoteTLS.Close(); err != nil {
+			p.log.Debugf("close remote TLS for %s: %v", hostname, err)
+		}
+	}()
+
+	negotiatedProto := clientTLS.ConnectionState().NegotiatedProtocol
+	p.log.Tracef("inspect: MITM established for %s (proto=%s)", hostname, negotiatedProto)
+
+	return p.inspectHTTP(ctx, clientTLS, remoteTLS, dst, sni, src, negotiatedProto)
+}
+
+// dialTLS connects to the destination with TLS, verifying the real server certificate.
+func (p *Proxy) dialTLS(ctx context.Context, dst netip.AddrPort, serverName string) (net.Conn, error) {
+	rawConn, err := p.dialTCP(ctx, dst)
+	if err != nil {
+		return nil, err
+	}
+
+	tlsConn := tls.Client(rawConn, &tls.Config{
+		ServerName: serverName,
+		NextProtos: []string{"h2", "http/1.1"},
+		MinVersion: tls.VersionTLS12,
+	})
+
+	if err := tlsConn.HandshakeContext(ctx); err != nil {
+		if closeErr := rawConn.Close(); closeErr != nil {
+			p.log.Debugf("close raw conn after TLS handshake failure: %v", closeErr)
+		}
+		return nil, fmt.Errorf("TLS handshake with %s: %w", serverName, err)
+	}
+
+	return tlsConn, nil
+}
+
+// protoFromALPN maps TLS ALPN protocol names to proxy ProtoType.
+// Falls back to ProtoHTTPS when no recognized ALPN is present.
+func protoFromALPN(alpn []string) ProtoType {
+	for _, p := range alpn {
+		switch p {
+		case "h2":
+			return ProtoH2
+		case "h3": // unlikely in TLS, but handle anyway
+			return ProtoH3
+		}
+	}
+	// No ALPN or only "http/1.1": treat as HTTPS
+	return ProtoHTTPS
+}
+
+// relay copies data bidirectionally between client and remote until one
+// side closes or the context is cancelled.
+func relay(ctx context.Context, client, remote net.Conn) error {
+	ctx, cancel := context.WithCancel(ctx)
+	defer cancel()
+
+	errCh := make(chan error, 2)
+
+	go func() {
+		_, err := io.Copy(remote, client)
+		cancel()
+		errCh <- err
+	}()
+
+	go func() {
+		_, err := io.Copy(client, remote)
+		cancel()
+		errCh <- err
+	}()
+
+	var firstErr error
+	for range 2 {
+		if err := <-errCh; err != nil && firstErr == nil {
+			if !isClosedErr(err) {
+				firstErr = err
+			}
+		}
+	}
+
+	return firstErr
+}
+
+// evaluateAction runs rule evaluation and resolves the effective action.
+// Pass empty path for connection-level (TLS), non-empty for request-level (HTTP).
+func (p *Proxy) evaluateAction(src netip.Addr, sni domain.Domain, dst netip.AddrPort, proto ProtoType, path string) Action {
+	return p.rules.Evaluate(src, sni, dst.Addr(), dst.Port(), proto, path)
+}
+
+// dialTCP dials the destination, blocking connections to loopback, link-local,
+// multicast, and WG overlay network addresses.
+func (p *Proxy) dialTCP(ctx context.Context, dst netip.AddrPort) (net.Conn, error) {
+	ip := dst.Addr().Unmap()
+	if err := p.validateDialTarget(ip); err != nil {
+		return nil, fmt.Errorf("dial %s: %w", dst, err)
+	}
+	return p.dialer.DialContext(ctx, "tcp", dst.String())
+}
+
+// validateDialTarget blocks destinations that should never be dialed by the proxy.
+// Mirrors the route validation in systemops.validateRoute.
+func (p *Proxy) validateDialTarget(addr netip.Addr) error {
+	switch {
+	case !addr.IsValid():
+		return fmt.Errorf("invalid address")
+	case addr.IsLoopback():
+		return fmt.Errorf("loopback address not allowed")
+	case addr.IsLinkLocalUnicast(), addr.IsLinkLocalMulticast(), addr.IsInterfaceLocalMulticast():
+		return fmt.Errorf("link-local address not allowed")
+	case addr.IsMulticast():
+		return fmt.Errorf("multicast address not allowed")
+	case p.wgNetwork.IsValid() && p.wgNetwork.Contains(addr):
+		return fmt.Errorf("overlay network address not allowed")
+	case p.localIPs != nil && p.localIPs.IsLocalIP(addr):
+		return fmt.Errorf("local address not allowed")
+	}
+	return nil
+}
+
+func isClosedErr(err error) bool {
+	if err == nil {
+		return false
+	}
+	return err == io.EOF ||
+		err == io.ErrClosedPipe ||
+		err == net.ErrClosed ||
+		err == context.Canceled
+}