feat(mgmt): add per-route scope enforcement on management HTTP port

CHANGELOG.md

@@ -8,6 +8,21 @@ adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ### Added
 
+- **Per-route scope enforcement on the management HTTP port.**
+  `WithMgmtControlAuth` is a new option that wraps the cluster-
+  mutating control endpoints (`POST /evict`, `POST /clear`,
+  `POST /trigger-expiration`) in a stricter auth gate than the
+  observability surface. The hypercache-server binary now wires
+  read-or-better on `/stats`/`/config`/`/cluster/*`/`/dist/*` and
+  admin-only on the control routes (see `cmd/hypercache-server/
+main.go`). `/health` is intentionally NOT auth-wrapped — k8s
+  liveness probes don't carry credentials, and a probe failure
+  cascades into a pod-restart loop. Also new: `httpauth.Policy.Verify`,
+  the "block-with-error" sibling of `Middleware()` that adapters
+  (like `WithMgmtAuth`/`WithMgmtControlAuth`) use when they own
+  their own next-handler dispatch. Existing `Middleware()` is now
+  thin sugar over `Verify() + c.Next()` so the auth logic lives
+  in exactly one place.
 - **`GET /v1/me` — resolved caller identity.** New scope-protected
   (`read`) route that reads the resolved `httpauth.Identity` from
   `c.Locals(httpauth.IdentityKey)` and returns

cmd/hypercache-server/main.go

@@ -266,9 +266,33 @@ func buildHyperCache(ctx context.Context, cfg envConfig, logger *slog.Logger) (*
 		)
 	}
 
+	// Phase C2: light up scope enforcement on the management port.
+	// /health stays public (k8s liveness probes carry no creds).
+	// Read-or-better is required for the observability surface
+	// (/stats, /config, /dist/*, /cluster/*); admin scope is
+	// required for the cluster-mutating control routes (/evict,
+	// /clear, /trigger-expiration). Closes a long-standing gap
+	// where the mgmt port was fully unauthenticated server-side
+	// while the monitor's proxy carried the only check.
+	//
+	// Closure captures cfg.AuthPolicy by value — Policy is value-
+	// semantic and safe for concurrent use after construction;
+	// see pkg/httpauth/policy.go.
+	policy := cfg.AuthPolicy
+	mgmtReadAuth := func(fiberCtx fiber.Ctx) error {
+		return policy.Verify(fiberCtx, httpauth.ScopeRead)
+	}
+	mgmtAdminAuth := func(fiberCtx fiber.Ctx) error {
+		return policy.Verify(fiberCtx, httpauth.ScopeAdmin)
+	}
+
 	hcCfg.HyperCacheOptions = append(
 		hcCfg.HyperCacheOptions,
-		hypercache.WithManagementHTTP[backend.DistMemory](cfg.MgmtAddr),
+		hypercache.WithManagementHTTP[backend.DistMemory](
+			cfg.MgmtAddr,
+			hypercache.WithMgmtAuth(mgmtReadAuth),
+			hypercache.WithMgmtControlAuth(mgmtAdminAuth),
+		),
 	)
 
 	hc, err := hypercache.New(ctx, hypercache.GetDefaultManager(), hcCfg)

cspell.config.yaml

@@ -77,6 +77,7 @@ words:
   - distroless
   - EDITMSG
   - elif
+  - Equalf
   - errcheck
   - errp
   - ewrap

management_http.go

@@ -18,14 +18,22 @@ type ManagementHTTPOption func(*ManagementHTTPServer)
 
 // ManagementHTTPServer holds Fiber app and settings.
 type ManagementHTTPServer struct {
-	addr             string
-	app              *fiber.App
-	readTimeout      time.Duration
-	writeTimeout     time.Duration
-	idleTimeout      time.Duration
-	bodyLimit        int
-	concurrency      int
-	authFunc         func(fiber.Ctx) error
+	addr         string
+	app          *fiber.App
+	readTimeout  time.Duration
+	writeTimeout time.Duration
+	idleTimeout  time.Duration
+	bodyLimit    int
+	concurrency  int
+	authFunc     func(fiber.Ctx) error
+	// controlAuthFunc is an optional stricter auth gate applied
+	// only to the cluster-mutating control endpoints (/evict,
+	// /clear, /trigger-expiration). When set, it runs INSTEAD OF
+	// authFunc on those routes — typically configured to require
+	// admin scope while authFunc requires read. When nil, the
+	// control routes fall back to authFunc, preserving the
+	// pre-Phase-C2 single-gate behavior.
+	controlAuthFunc  func(fiber.Ctx) error
 	ln               net.Listener
 	started          bool
 	listenerDeadline time.Duration
@@ -49,11 +57,28 @@ type ManagementHTTPServer struct {
 	serveErr atomic.Pointer[error]
 }
 
-// WithMgmtAuth sets an auth function (return error to block).
+// WithMgmtAuth sets an auth function applied to every authenticated
+// route on the management port (return error to block). /health is
+// exempt — k8s liveness probes do not carry credentials.
+//
+// Pair with WithMgmtControlAuth for finer scope on the cluster-
+// mutating endpoints (/evict, /clear, /trigger-expiration); without
+// it, those routes fall back to this same gate.
 func WithMgmtAuth(fn func(fiber.Ctx) error) ManagementHTTPOption {
 	return func(s *ManagementHTTPServer) { s.authFunc = fn }
 }
 
+// WithMgmtControlAuth sets a stricter auth function applied only to
+// the cluster-mutating control endpoints — /evict, /clear,
+// /trigger-expiration. Use this with httpauth.Policy.Verify(c,
+// httpauth.ScopeAdmin) so a token granted only read or write
+// scope cannot trigger destructive operations through the mgmt
+// port. When nil, control routes inherit authFunc's gate (the
+// pre-Phase-C2 single-gate behavior).
+func WithMgmtControlAuth(fn func(fiber.Ctx) error) ManagementHTTPOption {
+	return func(s *ManagementHTTPServer) { s.controlAuthFunc = fn }
+}
+
 // WithMgmtReadTimeout sets read timeout.
 func WithMgmtReadTimeout(d time.Duration) ManagementHTTPOption {
 	return func(s *ManagementHTTPServer) { s.readTimeout = d }
@@ -254,20 +279,42 @@ func (s *ManagementHTTPServer) Shutdown(ctx context.Context) error {
 // mountRoutes.
 func (s *ManagementHTTPServer) mountRoutes(hc managementCache) { // split into helpers to satisfy funlen
 	useAuth := s.wrapAuth
+	useControlAuth := s.wrapControlAuth
 	s.registerBasic(useAuth, hc)
 	s.registerDistributed(useAuth, hc)
 	s.registerCluster(useAuth, hc)
-	s.registerControl(useAuth, hc)
+	s.registerControl(useControlAuth, hc)
 }
 
 // wrapAuth returns an auth-wrapped handler if authFunc provided.
 func (s *ManagementHTTPServer) wrapAuth(handler fiber.Handler) fiber.Handler {
-	if s.authFunc == nil {
+	return wrapWithGate(s.authFunc, handler)
+}
+
+// wrapControlAuth returns a handler wrapped with the stricter
+// control-route auth when controlAuthFunc is set, otherwise it
+// falls back to wrapAuth. This preserves the pre-Phase-C2
+// single-gate behavior for operators who haven't opted into
+// admin-scope enforcement on the mgmt port.
+func (s *ManagementHTTPServer) wrapControlAuth(handler fiber.Handler) fiber.Handler {
+	if s.controlAuthFunc != nil {
+		return wrapWithGate(s.controlAuthFunc, handler)
+	}
+
+	return s.wrapAuth(handler)
+}
+
+// wrapWithGate applies an auth-gate function before invoking the
+// underlying handler. Nil gate is a passthrough — same shape as
+// before WithMgmtAuth was wired, used by deployments that haven't
+// configured any auth on the mgmt port.
+func wrapWithGate(gate func(fiber.Ctx) error, handler fiber.Handler) fiber.Handler {
+	if gate == nil {
 		return handler
 	}
 
 	return func(fiberCtx fiber.Ctx) error {
-		authErr := s.authFunc(fiberCtx)
+		authErr := gate(fiberCtx)
 		if authErr != nil {
 			return authErr
 		}
@@ -277,7 +324,12 @@ func (s *ManagementHTTPServer) wrapAuth(handler fiber.Handler) fiber.Handler {
 }
 
 func (s *ManagementHTTPServer) registerBasic(useAuth func(fiber.Handler) fiber.Handler, hc managementCache) {
-	s.app.Get("/health", useAuth(func(fiberCtx fiber.Ctx) error { return fiberCtx.SendString("ok") }))
+	// /health is intentionally NOT wrapped in useAuth — k8s
+	// liveness/readiness probes do not carry credentials, and
+	// a probe failure cascades into a pod-restart loop. Mirrors
+	// the client-API binary's `/healthz` exemption (see
+	// cmd/hypercache-server/main.go:registerClientRoutes).
+	s.app.Get("/health", func(fiberCtx fiber.Ctx) error { return fiberCtx.SendString("ok") })
 	s.app.Get("/stats", useAuth(func(fiberCtx fiber.Ctx) error { return fiberCtx.JSON(hc.GetStats()) }))
 	s.app.Get("/config", useAuth(func(fiberCtx fiber.Ctx) error {
 		cfg := map[string]any{

pkg/httpauth/policy.go

@@ -210,19 +210,48 @@ func (p Policy) Validate() error {
 // that want any-authenticated-caller semantics.
 func (p Policy) Middleware(required Scope) fiber.Handler {
 	return func(c fiber.Ctx) error {
-		identity, ok := p.resolve(c)
-		if !ok {
-			return c.SendStatus(fiber.StatusUnauthorized)
+		err := p.Verify(c, required)
+		if err != nil {
+			return err
 		}
 
-		if required != "" && !identity.HasScope(required) {
-			return c.SendStatus(fiber.StatusForbidden)
-		}
+		return c.Next()
+	}
+}
 
-		c.Locals(IdentityKey, identity)
+// Verify resolves credentials, asserts the required scope, and
+// stores the resolved Identity in c.Locals(IdentityKey). Returns
+// nil on success; on failure returns a *fiber.Error carrying
+// status 401 (no credentials matched) or 403 (credentials matched
+// but scope is missing). Fiber's default error handler emits the
+// canonical text body for the status code.
+//
+// Use Verify when integrating with code that owns its own next-
+// handler dispatch — e.g. ManagementHTTPServer.WithMgmtAuth and
+// WithMgmtControlAuth, which short-circuit on a non-nil return
+// from the gate function and never call the wrapped handler.
+// Middleware() is thin sugar over Verify() + Next() so the auth
+// logic lives in exactly one place.
+//
+// CRITICAL: do NOT switch to `c.SendStatus(...)` here. SendStatus
+// returns nil on success, which would silently fall through to
+// the wrapped handler in wrapWithGate-style adapters and the
+// downstream handler would write its own success status over the
+// 401 body. Returning a *fiber.Error keeps both Middleware and
+// the gate adapters fail-closed.
+func (p Policy) Verify(c fiber.Ctx, required Scope) error {
+	identity, ok := p.resolve(c)
+	if !ok {
+		return fiber.NewError(fiber.StatusUnauthorized)
+	}
 
-		return c.Next()
+	if required != "" && !identity.HasScope(required) {
+		return fiber.NewError(fiber.StatusForbidden)
 	}
+
+	c.Locals(IdentityKey, identity)
+
+	return nil
 }
 
 // resolve walks the credential resolution chain in priority order:

pkg/httpauth/policy_test.go

@@ -366,3 +366,114 @@ func TestPolicy_HasScope(t *testing.T) {
 		t.Errorf("Write should match Write")
 	}
 }
+
+// TestPolicy_Verify covers the public Verify() entry point — the
+// "block-with-error" sibling of Middleware that adapters
+// (ManagementHTTPServer.WithMgmtControlAuth, etc.) use when they
+// own their own next-handler dispatch. Same semantics as
+// Middleware: 401 on missing/invalid creds, 403 on wrong scope,
+// nil + Identity in Locals on success. The shared resolve() means
+// any future divergence between Middleware and Verify would be a
+// security bug; this test pins parity.
+func TestPolicy_Verify(t *testing.T) {
+	t.Parallel()
+
+	p := Policy{
+		Tokens: []TokenIdentity{
+			{ID: "ro", Token: "ro-token", Scopes: []Scope{ScopeRead}},
+			{ID: "admin", Token: "admin-token", Scopes: []Scope{ScopeRead, ScopeWrite, ScopeAdmin}},
+		},
+	}
+
+	cases := []struct {
+		name   string
+		header string
+		scope  Scope
+		want   int
+	}{
+		{"no creds → 401", "", ScopeRead, http.StatusUnauthorized},
+		{"bad bearer → 401", "Bearer wrong", ScopeRead, http.StatusUnauthorized},
+		{"read scope on read route → 200", "Bearer ro-token", ScopeRead, http.StatusOK},
+		{"read scope on admin route → 403", "Bearer ro-token", ScopeAdmin, http.StatusForbidden},
+		{"admin scope on admin route → 200", "Bearer admin-token", ScopeAdmin, http.StatusOK},
+		{"empty scope (any-authenticated) → 200 with creds", "Bearer ro-token", "", http.StatusOK},
+		{"empty scope still 401 without creds", "", "", http.StatusUnauthorized},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+
+			app := fiber.New()
+			// Mount Verify in the wrapAuth-style adapter shape:
+			// auth happens, then handler runs only on nil error.
+			// This is exactly how ManagementHTTPServer wires it.
+			scope := tc.scope
+			app.Get("/protected", func(c fiber.Ctx) error {
+				err := p.Verify(c, scope)
+				if err != nil {
+					return err
+				}
+
+				return c.SendString("ok")
+			})
+
+			got := doStatus(t, app, tc.header)
+			if got != tc.want {
+				t.Fatalf("status: got %d, want %d", got, tc.want)
+			}
+		})
+	}
+}
+
+// TestPolicy_Verify_StoresIdentityInLocals pins the side-effect
+// contract: a successful Verify populates IdentityKey before
+// returning. Adapters that read c.Locals(IdentityKey) — e.g. any
+// future audit-attribution handler on the mgmt port — depend on
+// it. Without this assertion a future refactor could regress
+// Verify into "scope-check only" silently.
+func TestPolicy_Verify_StoresIdentityInLocals(t *testing.T) {
+	t.Parallel()
+
+	p := Policy{
+		Tokens: []TokenIdentity{
+			{ID: "audit-target", Token: "tok", Scopes: []Scope{ScopeRead}},
+		},
+	}
+
+	app := fiber.New()
+	app.Get("/who", func(c fiber.Ctx) error {
+		err := p.Verify(c, ScopeRead)
+		if err != nil {
+			return err
+		}
+
+		v := c.Locals(IdentityKey)
+
+		id, ok := v.(Identity)
+		if !ok {
+			return c.Status(http.StatusInternalServerError).SendString("no identity")
+		}
+
+		return c.SendString(id.ID)
+	})
+
+	req := httptest.NewRequestWithContext(t.Context(), http.MethodGet, "/who", strings.NewReader(""))
+	req.Header.Set("Authorization", "Bearer tok")
+
+	resp, err := app.Test(req)
+	if err != nil {
+		t.Fatalf("app.Test: %v", err)
+	}
+
+	defer func() { _ = resp.Body.Close() }()
+
+	body := make([]byte, 64)
+
+	n, _ := resp.Body.Read(body)
+
+	got := string(body[:n])
+	if got != "audit-target" {
+		t.Fatalf("locals identity ID = %q, want %q", got, "audit-target")
+	}
+}