feat(breakglass): 恢复模式用邮件验证码证明管理员身份，发不出或验不过走带原因审计的 OVERRIDE (#13)

cmd/felis/breakglass.go

@@ -17,6 +17,7 @@ import (
 
 	"felis.lolicon.best/internal/api"
 	"felis.lolicon.best/internal/config"
+	"felis.lolicon.best/internal/platform"
 	"felis.lolicon.best/internal/store"
 
 	tea "github.com/charmbracelet/bubbletea"
@@ -33,13 +34,15 @@ import (
 //
 // Root is necessary but NOT sufficient for accountability: root is machine
 // authority, not a human identity, so the console additionally captures WHO is
-// breaking the glass. When a staff account already exists it asks the operator to
-// authenticate as an existing admin (the verified identity is the accountable
-// actor); when none exists yet it bootstraps the first Owner from the typed
-// credential and attributes the act to the OS user. The audit row records the
-// difference. This attribution is best-effort, not tamper-proof — whoever runs
-// this is root and can edit Postgres directly — but it produces an honest trail
-// for an honest operator, which is the point.
+// breaking the glass. When a staff account already exists the operator names one
+// and types the one-time code the console mails to its verified address
+// (breakglass_otp.go); that account is then the accountable actor. When no code can
+// be sent or proven, the typed OVERRIDE proceeds as the OS user and the audit row
+// says why. When no staff account exists yet it bootstraps the first Owner and
+// attributes the act to the OS user. The audit row records which of these
+// happened. This attribution is best-effort, not tamper-proof — whoever runs this
+// is root and can edit Postgres directly — but it produces an honest trail for an
+// honest operator, which is the point.
 //
 // When a staff account already exists the console opens on a thin top-level menu
 // (menuModel) so that operations are peers, not tails of one wizard. Two account
@@ -62,7 +65,7 @@ import (
 // suspension for the interactive `cloudflared tunnel login` browser consent.
 
 // breakGlassOverrideToken is the literal an operator must type to proceed when no
-// admin credential could be verified. Requiring an explicit, deliberate word (not a
+// admin could be verified by a mailed code. Requiring an explicit, deliberate word (not a
 // bare Enter) keeps the unverified root override from happening by reflex.
 const breakGlassOverrideToken = "OVERRIDE"
 
@@ -142,7 +145,7 @@ func cmdBreakGlass(args []string, stdout, stderr io.Writer) int {
 	repo := api.NewPGRepo(drv.DB())
 
 	// Decide bootstrap (no admin yet → typed credential mints the first Owner) vs
-	// recovery (an admin exists → the operator must authenticate as one) BEFORE the
+	// recovery (an admin exists → the operator proves one with a mailed code) BEFORE the
 	// alt-screen TUI takes over, so a database fault surfaces as a plain error.
 	adminExists, err := repo.AdminExists(ctx)
 	if err != nil {
@@ -150,7 +153,12 @@ func cmdBreakGlass(args []string, stdout, stderr io.Writer) int {
 		return 1
 	}
 
-	res, err := runBreakGlassTUI(ctx, repo, cfg.Database.URL, cfg.Server.RootDomain, cfg.Auth.AdminHostname, cfg.Auth.PanelHostname, cfg.Auth.AccessJWTAud, cfg.K8s.Namespace, accountableOSUser(), adminExists)
+	// Recovery mails its code through [smtp]; the relay is opened only if a code is
+	// asked for.
+	host, _ := os.Hostname()
+	recovery := recoveryConfig{open: hostRecoveryMailer(cfg.SMTP, platform.DefaultControlNamespace), host: host}
+
+	res, err := runBreakGlassTUI(ctx, repo, cfg.Database.URL, cfg.Server.RootDomain, cfg.Auth.AdminHostname, cfg.Auth.PanelHostname, cfg.Auth.AccessJWTAud, cfg.K8s.Namespace, accountableOSUser(), adminExists, recovery)
 	if err != nil {
 		fmt.Fprintf(stderr, "felis breakGlass: %v\n", err)
 		return 1
@@ -174,6 +182,9 @@ func cmdBreakGlass(args []string, stdout, stderr io.Writer) int {
 			fmt.Fprintf(stdout, "\nfelis breakGlass: Owner account %q provisioned; local session sign-in is ENABLED.\n", res.username)
 		}
 		fmt.Fprintf(stdout, "Recorded as %q (mode: %s, os user: %s).\n", res.accountable, res.mode, res.osUser)
+		if res.mode == "root_override" {
+			fmt.Fprintln(stdout, "No admin was proven by an email code; the audit row records this run as an unverified root override and why.")
+		}
 		if res.setupTokenURL != "" {
 			fmt.Fprintf(stdout, "One-time setup URL (opens a lockdown session to verify email / enroll passkey):\n\n    %s\n\n", res.setupTokenURL)
 		}
@@ -258,29 +269,6 @@ func newOwnerID() string {
 	return "usr-" + hex.EncodeToString(b[:])
 }
 
-// authenticateAdmin resolves a typed admin username for recovery-mode attribution.
-// Password verification is gone (passwordless design); Phase 3 replaces this with
-// email-OTP recovery. For now it confirms the named admin exists.
-func authenticateAdmin(ctx context.Context, s ownerStore, username string) (matched string, ok bool, err error) {
-	username = strings.TrimSpace(username)
-	if username == "" {
-		return "", false, nil
-	}
-	u, err := s.UserByUsername(ctx, username)
-	if errors.Is(err, api.ErrNotFound) {
-		return "", false, nil
-	}
-	if err != nil {
-		return "", false, err
-	}
-	// Staff means admin OR owner: recovery attribution must accept the Owner (the
-	// primary break-glass identity), not just plain admins.
-	if u.Role != "admin" && u.Role != "owner" {
-		return "", false, nil
-	}
-	return u.Username, true, nil
-}
-
 // provisionOwner mints or resets the single Owner account direct-to-Postgres,
 // passwordless. The account is role=owner with no password — the Owner completes
 // passwordless login setup via the web setup-token flow after `felis setup`.
@@ -371,6 +359,10 @@ type breakGlassOp struct {
 	ownerUsername  string
 	ownerEmail     string
 	attemptedAdmin string // recovery / override: the admin username the operator typed
+	verifiedBy     string // recovery: how the admin was proven (verifiedByEmailOTP)
+	codeSentTo     string // recovery: the address the proving code went to
+	otpSkipped     string // root_override: why no code proved an admin (otpSkip*)
+	otpSkipDetail  string // root_override: what failed, when something did
 }
 
 // breakGlassOutcome is what performBreakGlass reports back to the TUI.
@@ -494,16 +486,7 @@ func auditSetupMCBind(ctx context.Context, s ownerStore, osUser, mcUUID, authSou
 // does not fail the recovery if this write fails — and intentionally honest: it
 // records attribution, it does not prove it (a malicious root can edit the row).
 func auditBreakGlass(ctx context.Context, s ownerStore, op breakGlassOp) error {
-	payload := map[string]any{
-		"mode":     op.mode,
-		"owner":    op.ownerUsername,
-		"os_user":  op.osUser,
-		"verified": op.mode == "recovery",
-	}
-	if op.attemptedAdmin != "" {
-		payload["admin_account"] = op.attemptedAdmin
-	}
-	blob, err := json.Marshal(payload)
+	blob, err := json.Marshal(breakGlassPayload(op, "owner"))
 	if err != nil {
 		return err
 	}
@@ -515,6 +498,33 @@ func auditBreakGlass(ctx context.Context, s ownerStore, op breakGlassOp) error {
 	})
 }
 
+// breakGlassPayload is the who/how both account audits carry, with the account the
+// run wrote under subjectKey. verified is true only for a run a mailed code proved;
+// such a run names the address the code went to, and an override names why no code
+// proved anyone.
+func breakGlassPayload(op breakGlassOp, subjectKey string) map[string]any {
+	payload := map[string]any{
+		"mode":     op.mode,
+		subjectKey: op.ownerUsername,
+		"os_user":  op.osUser,
+		"verified": op.verifiedBy != "",
+	}
+	if op.attemptedAdmin != "" {
+		payload["admin_account"] = op.attemptedAdmin
+	}
+	if op.verifiedBy != "" {
+		payload["verified_by"] = op.verifiedBy
+		payload["code_sent_to"] = op.codeSentTo
+	}
+	if op.otpSkipped != "" {
+		payload["otp_skipped"] = op.otpSkipped
+		if op.otpSkipDetail != "" {
+			payload["otp_skip_detail"] = op.otpSkipDetail
+		}
+	}
+	return payload
+}
+
 // performAddOperator mints a NEW Operator account and records a best-effort
 // accountability row. It mirrors performBreakGlass — passwordless — with two
 // deliberate differences. (1) It provisions insert-only (provisionOperator), so it
@@ -538,16 +548,7 @@ func performAddOperator(ctx context.Context, s ownerStore, op breakGlassOp) (bre
 // break_glass.operator_create action, naming the new account under an "operator" key
 // rather than "owner".
 func auditAddOperator(ctx context.Context, s ownerStore, op breakGlassOp) error {
-	payload := map[string]any{
-		"mode":     op.mode,
-		"operator": op.ownerUsername,
-		"os_user":  op.osUser,
-		"verified": op.mode == "recovery",
-	}
-	if op.attemptedAdmin != "" {
-		payload["admin_account"] = op.attemptedAdmin
-	}
-	blob, err := json.Marshal(payload)
+	blob, err := json.Marshal(breakGlassPayload(op, "operator"))
 	if err != nil {
 		return err
 	}
@@ -626,16 +627,19 @@ const (
 	cloudflareAPITokenDocsURL        = "https://developers.cloudflare.com/fundamentals/api/how-to/account-owned-token-template/"
 )
 
-func runBreakGlassTUI(ctx context.Context, s ownerStore, dbURL, rootDomain, adminHostname, panelHostname, accessAud, namespace, osUser string, adminExists bool) (breakGlassResult, error) {
-	return runConsoleTUI(ctx, s, dbURL, rootDomain, adminHostname, panelHostname, accessAud, namespace, osUser, adminExists, consoleModeBreakGlass)
+func runBreakGlassTUI(ctx context.Context, s ownerStore, dbURL, rootDomain, adminHostname, panelHostname, accessAud, namespace, osUser string, adminExists bool, recovery recoveryConfig) (breakGlassResult, error) {
+	return runConsoleTUI(ctx, s, dbURL, rootDomain, adminHostname, panelHostname, accessAud, namespace, osUser, adminExists, consoleModeBreakGlass, recovery)
 }
 
+// runSetupTUI never reaches recovery: setup with a staff account present lands on
+// the status screen, so it has no relay to hand over.
 func runSetupTUI(ctx context.Context, s ownerStore, dbURL, rootDomain, adminHostname, panelHostname, accessAud, namespace, osUser string, adminExists bool) (breakGlassResult, error) {
-	return runConsoleTUI(ctx, s, dbURL, rootDomain, adminHostname, panelHostname, accessAud, namespace, osUser, adminExists, consoleModeSetup)
+	return runConsoleTUI(ctx, s, dbURL, rootDomain, adminHostname, panelHostname, accessAud, namespace, osUser, adminExists, consoleModeSetup, recoveryConfig{})
 }
 
-func runConsoleTUI(ctx context.Context, s ownerStore, dbURL, rootDomain, adminHostname, panelHostname, accessAud, namespace, osUser string, adminExists bool, mode consoleMode) (breakGlassResult, error) {
+func runConsoleTUI(ctx context.Context, s ownerStore, dbURL, rootDomain, adminHostname, panelHostname, accessAud, namespace, osUser string, adminExists bool, mode consoleMode, recovery recoveryConfig) (breakGlassResult, error) {
 	rm := newRootModel(ctx, s, dbURL, rootDomain, adminHostname, panelHostname, accessAud, namespace, osUser, adminExists, mode)
+	rm.recovery = recovery
 	final, err := tea.NewProgram(rm, tea.WithAltScreen()).Run()
 	if err != nil {
 		return breakGlassResult{}, err

cmd/felis/breakglass_otp.go

@@ -0,0 +1,252 @@
+package main
+
+import (
+	"context"
+	"crypto/rand"
+	"crypto/subtle"
+	"errors"
+	"fmt"
+	"math/big"
+	"os"
+	"strings"
+	"time"
+
+	"felis.lolicon.best/internal/api"
+	"felis.lolicon.best/internal/config"
+	"felis.lolicon.best/internal/platform"
+)
+
+// Recovery mode proves who is breaking the glass (#13). Naming a staff account is
+// where it starts: the console then mails a one-time code to that account's verified
+// address, and only that code, typed within recoveryCodeTTL, makes the run a
+// recovery attributed to the account. Every other ending — no such account, no
+// verified address, no relay, a send that fails, a wrong or late code, or the
+// operator giving up on the mail — leads to the typed OVERRIDE, which the audit row
+// records as an unverified root_override together with the reason (otp_skipped).
+// The code goes through the same [smtp] relay as the panel's login codes, so with
+// that relay down recovery still works, as an override that says why.
+
+const (
+	recoveryCodeTTL      = 10 * time.Minute
+	recoveryCodeAttempts = 5
+)
+
+// The reasons a run fell back to the override, recorded as otp_skipped.
+const (
+	otpSkipUnknownAdmin    = "unknown_admin"
+	otpSkipNoVerifiedEmail = "no_verified_email"
+	otpSkipNoRelay         = "no_relay"
+	otpSkipSendFailed      = "send_failed"
+	otpSkipCodeExpired     = "code_expired"
+	otpSkipCodeRejected    = "code_rejected"
+	otpSkipByOperator      = "operator_skipped"
+)
+
+// verifiedByEmailOTP is the audit's verified_by for a recovery the mailed code proved.
+const verifiedByEmailOTP = "email_otp"
+
+// recoveryMailer is the one relay call a recovery code needs; *mail.SMTP has it.
+type recoveryMailer interface {
+	SendNotice(ctx context.Context, email, subject, body string) error
+}
+
+// recoveryConfig is what the console needs to mail a recovery code. open resolves
+// the relay only when a code is about to go out, so a console used to halt a server
+// never touches [smtp] or the cluster; its error says why no relay is available.
+// host names this machine in the mail. The zero value has no relay.
+type recoveryConfig struct {
+	open func(ctx context.Context) (recoveryMailer, error)
+	host string
+	now  func() time.Time
+}
+
+func (r recoveryConfig) clock() time.Time {
+	if r.now != nil {
+		return r.now()
+	}
+	return time.Now()
+}
+
+// recoveryCode is one mailed code: its value, when it stops working, and how many
+// wrong codes were typed against it.
+type recoveryCode struct {
+	value    string
+	expires  time.Time
+	failures int
+}
+
+func newRecoveryCode(now time.Time) (*recoveryCode, error) {
+	n, err := rand.Int(rand.Reader, big.NewInt(1_000_000))
+	if err != nil {
+		return nil, fmt.Errorf("generate recovery code: %w", err)
+	}
+	return &recoveryCode{value: fmt.Sprintf("%06d", n.Int64()), expires: now.Add(recoveryCodeTTL)}, nil
+}
+
+type codeVerdict int
+
+const (
+	codeAccepted codeVerdict = iota
+	codeWrong
+	codeExpired
+	codeExhausted
+)
+
+// check compares a typed code in constant time. Each wrong code counts; the one
+// that reaches recoveryCodeAttempts exhausts the code, which then accepts nothing,
+// and neither does an expired one.
+func (c *recoveryCode) check(typed string, now time.Time) codeVerdict {
+	if c.failures >= recoveryCodeAttempts {
+		return codeExhausted
+	}
+	if !now.Before(c.expires) {
+		return codeExpired
+	}
+	if subtle.ConstantTimeCompare([]byte(strings.TrimSpace(typed)), []byte(c.value)) == 1 {
+		return codeAccepted
+	}
+	c.failures++
+	if c.failures >= recoveryCodeAttempts {
+		return codeExhausted
+	}
+	return codeWrong
+}
+
+func (c *recoveryCode) attemptsLeft() int { return recoveryCodeAttempts - c.failures }
+
+// recoveryStart is where naming an admin led: a code on its way to that admin, or
+// the reason the run has to fall back to the override.
+type recoveryStart struct {
+	admin  *api.StaffUser // the named staff account; nil when none matched
+	code   *recoveryCode  // set when the code went out
+	skip   string         // otpSkip* when it did not
+	detail string         // what failed, for the override screen and the audit row
+}
+
+// resolveAdmin loads the staff account (admin or owner) a typed username names, or
+// nil when there is none.
+func resolveAdmin(ctx context.Context, s ownerStore, username string) (*api.StaffUser, error) {
+	username = strings.TrimSpace(username)
+	if username == "" {
+		return nil, nil
+	}
+	u, err := s.UserByUsername(ctx, username)
+	if errors.Is(err, api.ErrNotFound) {
+		return nil, nil
+	}
+	if err != nil {
+		return nil, err
+	}
+	// Staff means admin or owner: the Owner is the primary break-glass identity.
+	if u.Role != "admin" && u.Role != "owner" {
+		return nil, nil
+	}
+	return u, nil
+}
+
+// beginRecovery resolves the named admin and mails it a recovery code. Only a
+// datastore or entropy fault is an error; every other way the code cannot go out is
+// a recoveryStart with skip set.
+func beginRecovery(ctx context.Context, s ownerStore, rc recoveryConfig, username, osUser string, op bgOperation) (recoveryStart, error) {
+	admin, err := resolveAdmin(ctx, s, username)
+	if err != nil {
+		return recoveryStart{}, err
+	}
+	if admin == nil {
+		return recoveryStart{skip: otpSkipUnknownAdmin}, nil
+	}
+	st := recoveryStart{admin: admin}
+	// An address nobody ever proved vouches for nobody.
+	email := strings.TrimSpace(admin.Email)
+	if email == "" || !admin.EmailVerified {
+		st.skip = otpSkipNoVerifiedEmail
+		return st, nil
+	}
+	if rc.open == nil {
+		st.skip, st.detail = otpSkipNoRelay, "this console has no mail relay"
+		return st, nil
+	}
+	relay, err := rc.open(ctx)
+	if err != nil {
+		st.skip, st.detail = otpSkipNoRelay, err.Error()
+		return st, nil
+	}
+	code, err := newRecoveryCode(rc.clock())
+	if err != nil {
+		return recoveryStart{}, err
+	}
+	sendCtx, cancel := context.WithTimeout(ctx, 30*time.Second)
+	defer cancel()
+	subject, body := recoveryMail(code.value, rc.host, osUser, admin.Username, op)
+	if err := relay.SendNotice(sendCtx, email, subject, body); err != nil {
+		st.skip, st.detail = otpSkipSendFailed, err.Error()
+		return st, nil
+	}
+	st.code = code
+	return st, nil
+}
+
+// recoveryMail words the code mail. It says where, by whom and for what the console
+// was opened, so an admin who did not ask for it learns that root on that machine is
+// in someone else's hands.
+func recoveryMail(code, host, osUser, admin string, op bgOperation) (subject, body string) {
+	what, whatZH := "reset the Owner account", "重置 Owner 账号"
+	if op == bgAddOperator {
+		what, whatZH = "add an Operator account", "添加 Operator 账号"
+	}
+	if host == "" {
+		host = "the Felis host"
+	}
+	minutes := int(recoveryCodeTTL / time.Minute)
+	subject = "Felis break-glass recovery code / 紧急恢复验证码"
+	body = fmt.Sprintf(`Someone with root on %[1]s (OS user %[2]s) opened felis breakGlass and named your staff account %[3]q to %[4]s.
+
+Recovery code: %[6]s
+It works for %[7]d minutes.
+
+If this was not you, root on that machine is in someone else's hands: change its credentials and read the audit log for break_glass entries.
+
+有人在 %[1]s 上以 root 身份（系统用户 %[2]s）打开了 felis breakGlass，指名你的管理员账号 %[3]q 来%[5]s。
+
+恢复验证码：%[6]s
+%[7]d 分钟内有效。
+
+如果不是你本人，这台机器的 root 已落入他人之手：请更换它的凭据，并查看审计日志中的 break_glass 记录。
+`, host, osUser, admin, what, whatZH, code, minutes)
+	return subject, body
+}
+
+// maskEmail keeps the first character of the local part and the domain, enough for
+// the operator to recognise the address without putting it on screen whole.
+func maskEmail(email string) string {
+	at := strings.LastIndex(email, "@")
+	if at <= 0 {
+		return "***"
+	}
+	return email[:1] + strings.Repeat("*", max(at-1, 3)) + email[at:]
+}
+
+// hostRecoveryMailer opens the [smtp] relay from the host the way the watchdog does:
+// the password is the env var password_ref names when that is set, else the
+// felis-smtp Secret, whose absence means a relay without AUTH.
+func hostRecoveryMailer(c config.SMTPConfig, controlNS string) func(context.Context) (recoveryMailer, error) {
+	return func(ctx context.Context) (recoveryMailer, error) {
+		if strings.TrimSpace(c.Host) == "" {
+			return nil, errors.New("[smtp] is not configured in felis.toml")
+		}
+		if ref := c.PasswordRef; ref != "" && os.Getenv(ref) != "" {
+			return smtpRelay(c, os.Getenv(ref)), nil
+		}
+		cl, err := buildSystemServerClient()
+		if err != nil {
+			return nil, fmt.Errorf("reach the cluster for the relay password: %w", err)
+		}
+		ctx, cancel := context.WithTimeout(ctx, 15*time.Second)
+		defer cancel()
+		password, err := smtpSecretPassword(ctx, cl, controlNS)
+		if err != nil {
+			return nil, fmt.Errorf("read the relay password from %s/%s: %w", controlNS, platform.SMTPSecretName, err)
+		}
+		return smtpRelay(c, password), nil
+	}
+}

cmd/felis/breakglass_otp_test.go

@@ -0,0 +1,498 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"strings"
+	"testing"
+	"time"
+
+	"felis.lolicon.best/internal/api"
+	"felis.lolicon.best/internal/config"
+	"felis.lolicon.best/internal/mail"
+	"felis.lolicon.best/internal/platform"
+
+	tea "github.com/charmbracelet/bubbletea"
+	corev1 "k8s.io/api/core/v1"
+	apierrors "k8s.io/apimachinery/pkg/api/errors"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/runtime/schema"
+	"sigs.k8s.io/controller-runtime/pkg/client"
+	"sigs.k8s.io/controller-runtime/pkg/client/fake"
+	"sigs.k8s.io/controller-runtime/pkg/client/interceptor"
+)
+
+// These tests cover the recovery proof (#13): the mailed code's rules, where
+// naming an admin leads, what the mail says, how the console walks from a name to
+// a proven (or overridden) run, and what the audit row then records. No mail
+// leaves the process: the relay is a fake that keeps what it was handed.
+
+type sentMail struct{ to, subject, body string }
+
+type fakeRelay struct {
+	sent []sentMail
+	err  error
+}
+
+func (r *fakeRelay) SendNotice(_ context.Context, to, subject, body string) error {
+	if r.err != nil {
+		return r.err
+	}
+	r.sent = append(r.sent, sentMail{to, subject, body})
+	return nil
+}
+
+// sentCode pulls the code out of the one mail the relay carried.
+func (r *fakeRelay) sentCode(t *testing.T) string {
+	t.Helper()
+	if len(r.sent) != 1 {
+		t.Fatalf("relay carried %d mails, want 1", len(r.sent))
+	}
+	_, after, ok := strings.Cut(r.sent[0].body, "Recovery code: ")
+	if !ok || len(after) < 6 {
+		t.Fatalf("mail carries no recovery code:\n%s", r.sent[0].body)
+	}
+	return after[:6]
+}
+
+func relayConfig(r *fakeRelay, now func() time.Time) recoveryConfig {
+	return recoveryConfig{
+		open: func(context.Context) (recoveryMailer, error) { return r, nil },
+		host: "felis-host-1",
+		now:  now,
+	}
+}
+
+func verifiedAdmin(username, email string) *api.StaffUser {
+	return &api.StaffUser{ID: "usr-" + username, Username: username, Role: "owner", Email: email, EmailVerified: true}
+}
+
+func TestRecoveryCodeRules(t *testing.T) {
+	t0 := time.Date(2026, 9, 26, 8, 0, 0, 0, time.UTC)
+
+	t.Run("a fresh code is six digits and lives for the TTL", func(t *testing.T) {
+		seen := map[string]bool{}
+		for i := 0; i < 20; i++ {
+			c, err := newRecoveryCode(t0)
+			if err != nil {
+				t.Fatal(err)
+			}
+			if !isRecoveryCodeShape(c.value) {
+				t.Fatalf("code %q is not six digits", c.value)
+			}
+			if !c.expires.Equal(t0.Add(recoveryCodeTTL)) {
+				t.Fatalf("expires = %v, want %v", c.expires, t0.Add(recoveryCodeTTL))
+			}
+			seen[c.value] = true
+		}
+		if len(seen) < 2 {
+			t.Error("twenty codes were all the same")
+		}
+	})
+
+	t.Run("the right code is accepted, surrounding space ignored", func(t *testing.T) {
+		c := &recoveryCode{value: "042917", expires: t0.Add(recoveryCodeTTL)}
+		if v := c.check(" 042917 ", t0); v != codeAccepted {
+			t.Errorf("check = %v, want accepted", v)
+		}
+	})
+
+	t.Run("wrong codes count down and the last one exhausts it", func(t *testing.T) {
+		c := &recoveryCode{value: "042917", expires: t0.Add(recoveryCodeTTL)}
+		for i := 1; i < recoveryCodeAttempts; i++ {
+			if v := c.check("000000", t0); v != codeWrong {
+				t.Fatalf("wrong code %d: check = %v, want wrong", i, v)
+			}
+			if left := c.attemptsLeft(); left != recoveryCodeAttempts-i {
+				t.Fatalf("after %d wrong codes attemptsLeft = %d, want %d", i, left, recoveryCodeAttempts-i)
+			}
+		}
+		if v := c.check("000000", t0); v != codeExhausted {
+			t.Fatalf("wrong code %d: check = %v, want exhausted", recoveryCodeAttempts, v)
+		}
+		if v := c.check("042917", t0); v != codeExhausted {
+			t.Errorf("the right code after exhaustion: check = %v, want exhausted", v)
+		}
+	})
+
+	t.Run("an expired code accepts nothing", func(t *testing.T) {
+		c := &recoveryCode{value: "042917", expires: t0.Add(recoveryCodeTTL)}
+		if v := c.check("042917", t0.Add(recoveryCodeTTL-time.Second)); v != codeAccepted {
+			t.Fatalf("a second before expiry: check = %v, want accepted", v)
+		}
+		c = &recoveryCode{value: "042917", expires: t0.Add(recoveryCodeTTL)}
+		if v := c.check("042917", t0.Add(recoveryCodeTTL)); v != codeExpired {
+			t.Errorf("at expiry: check = %v, want expired", v)
+		}
+	})
+}
+
+func TestBeginRecovery(t *testing.T) {
+	ctx := context.Background()
+	t0 := time.Date(2026, 9, 26, 8, 0, 0, 0, time.UTC)
+	clock := func() time.Time { return t0 }
+
+	t.Run("mails a code to the named admin's verified address", func(t *testing.T) {
+		f := &fakeOwnerStore{users: map[string]*api.StaffUser{"root": verifiedAdmin("root", "[email protected]")}}
+		r := &fakeRelay{}
+		st, err := beginRecovery(ctx, f, relayConfig(r, clock), "root", "alice", bgAddOperator)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if st.code == nil || st.skip != "" {
+			t.Fatalf("start = %+v, want a code and no skip", st)
+		}
+		if st.admin == nil || st.admin.Username != "root" {
+			t.Fatalf("start.admin = %+v, want root", st.admin)
+		}
+		if got := r.sentCode(t); got != st.code.value {
+			t.Errorf("mailed code %q, want the code the console checks (%q)", got, st.code.value)
+		}
+		m := r.sent[0]
+		if m.to != "[email protected]" {
+			t.Errorf("mail went to %q, want [email protected]", m.to)
+		}
+		for _, want := range []string{"felis-host-1", "OS user alice", `"root"`, "add an Operator account", "添加 Operator 账号", "10 minutes"} {
+			if !strings.Contains(m.body, want) {
+				t.Errorf("mail body lacks %q:\n%s", want, m.body)
+			}
+		}
+		if !st.code.expires.Equal(t0.Add(recoveryCodeTTL)) {
+			t.Errorf("code expires %v, want %v", st.code.expires, t0.Add(recoveryCodeTTL))
+		}
+	})
+
+	t.Run("the Owner reset is named as such", func(t *testing.T) {
+		f := &fakeOwnerStore{users: map[string]*api.StaffUser{"root": verifiedAdmin("root", "[email protected]")}}
+		r := &fakeRelay{}
+		if _, err := beginRecovery(ctx, f, relayConfig(r, clock), "root", "alice", bgProvisionOwner); err != nil {
+			t.Fatal(err)
+		}
+		if body := r.sent[0].body; !strings.Contains(body, "reset the Owner account") || !strings.Contains(body, "重置 Owner 账号") {
+			t.Errorf("mail body does not name the Owner reset:\n%s", body)
+		}
+	})
+
+	// Each way the code cannot go out ends in a skip reason and no mail.
+	unverified := verifiedAdmin("root", "[email protected]")
+	unverified.EmailVerified = false
+	noEmail := verifiedAdmin("root", "")
+	cases := []struct {
+		name       string
+		user       *api.StaffUser
+		rc         func(r *fakeRelay) recoveryConfig
+		skip       string
+		detail     string
+		wantAdmin  bool
+		relayError error
+	}{
+		{name: "unknown admin", user: nil, rc: func(r *fakeRelay) recoveryConfig { return relayConfig(r, clock) }, skip: otpSkipUnknownAdmin},
+		{name: "unverified address", user: unverified, rc: func(r *fakeRelay) recoveryConfig { return relayConfig(r, clock) }, skip: otpSkipNoVerifiedEmail, wantAdmin: true},
+		{name: "no address", user: noEmail, rc: func(r *fakeRelay) recoveryConfig { return relayConfig(r, clock) }, skip: otpSkipNoVerifiedEmail, wantAdmin: true},
+		{name: "no relay wired", user: verifiedAdmin("root", "[email protected]"), rc: func(*fakeRelay) recoveryConfig { return recoveryConfig{} }, skip: otpSkipNoRelay, detail: "this console has no mail relay", wantAdmin: true},
+		{name: "relay cannot open", user: verifiedAdmin("root", "[email protected]"), rc: func(*fakeRelay) recoveryConfig {
+			return recoveryConfig{open: func(context.Context) (recoveryMailer, error) {
+				return nil, errors.New("[smtp] is not configured in felis.toml")
+			}}
+		}, skip: otpSkipNoRelay, detail: "[smtp] is not configured in felis.toml", wantAdmin: true},
+		{name: "send fails", user: verifiedAdmin("root", "[email protected]"), rc: func(r *fakeRelay) recoveryConfig { return relayConfig(r, clock) },
+			skip: otpSkipSendFailed, detail: "554 relay refused", wantAdmin: true, relayError: errors.New("554 relay refused")},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			f := &fakeOwnerStore{users: map[string]*api.StaffUser{}}
+			if tc.user != nil {
+				f.users["root"] = tc.user
+			}
+			r := &fakeRelay{err: tc.relayError}
+			st, err := beginRecovery(ctx, f, tc.rc(r), "root", "alice", bgProvisionOwner)
+			if err != nil {
+				t.Fatal(err)
+			}
+			if st.code != nil || st.skip != tc.skip || st.detail != tc.detail {
+				t.Errorf("start = {code:%v skip:%q detail:%q}, want no code, skip %q, detail %q", st.code, st.skip, st.detail, tc.skip, tc.detail)
+			}
+			if (st.admin != nil) != tc.wantAdmin {
+				t.Errorf("start.admin = %+v, want present=%v", st.admin, tc.wantAdmin)
+			}
+			if len(r.sent) != 0 {
+				t.Errorf("relay carried %d mails, want none", len(r.sent))
+			}
+		})
+	}
+
+	t.Run("a datastore fault is an error", func(t *testing.T) {
+		f := &fakeOwnerStore{userErr: errors.New("db down")}
+		if _, err := beginRecovery(ctx, f, relayConfig(&fakeRelay{}, clock), "root", "alice", bgProvisionOwner); err == nil {
+			t.Fatal("want the store fault")
+		}
+	})
+}
+
+func TestMaskEmail(t *testing.T) {
+	for in, want := range map[string]string{
+		"[email protected]": "a****@example.com",
+		"[email protected]":    "a***@example.com",
+		"@example.com":      "***",
+		"nonsense":          "***",
+	} {
+		if got := maskEmail(in); got != want {
+			t.Errorf("maskEmail(%q) = %q, want %q", in, got, want)
+		}
+	}
+}
+
+func TestHostRecoveryMailer(t *testing.T) {
+	ctx := context.Background()
+
+	t.Run("no [smtp] host is no relay", func(t *testing.T) {
+		_, err := hostRecoveryMailer(config.SMTPConfig{}, "felis")(ctx)
+		if err == nil || !strings.Contains(err.Error(), "[smtp]") {
+			t.Fatalf("err = %v, want it to name [smtp]", err)
+		}
+	})
+
+	t.Run("the password_ref env var supplies the password", func(t *testing.T) {
+		t.Setenv("FELIS_TEST_RELAY_PW", "from-env")
+		off := false
+		c := config.SMTPConfig{Host: "mail.example.com", Port: 2525, From: "[email protected]", Username: "felis", PasswordRef: "FELIS_TEST_RELAY_PW", RequireTLS: &off}
+		got, err := hostRecoveryMailer(c, "felis")(ctx)
+		if err != nil {
+			t.Fatal(err)
+		}
+		relay, ok := got.(*mail.SMTP)
+		if !ok {
+			t.Fatalf("relay is %T, want *mail.SMTP", got)
+		}
+		if relay.Password != "from-env" || relay.Host != "mail.example.com" || relay.Port != 2525 || relay.From != "[email protected]" || relay.Username != "felis" || relay.RequireTLS {
+			t.Errorf("relay = %+v, want the [smtp] fields with the env password and TLS as configured", *relay)
+		}
+	})
+}
+
+func TestSMTPSecretPassword(t *testing.T) {
+	ctx := context.Background()
+	scheme := haltScheme(t)
+
+	t.Run("reads the password key", func(t *testing.T) {
+		cl := fake.NewClientBuilder().WithScheme(scheme).WithObjects(&corev1.Secret{
+			ObjectMeta: metav1.ObjectMeta{Namespace: "felis", Name: platform.SMTPSecretName},
+			Data:       map[string][]byte{platform.SMTPSecretPasswordKey: []byte("s3cret")},
+		}).Build()
+		if pw, err := smtpSecretPassword(ctx, cl, "felis"); err != nil || pw != "s3cret" {
+			t.Errorf("smtpSecretPassword = (%q, %v), want (s3cret, nil)", pw, err)
+		}
+	})
+
+	t.Run("a missing Secret is a relay without AUTH", func(t *testing.T) {
+		cl := fake.NewClientBuilder().WithScheme(scheme).Build()
+		if pw, err := smtpSecretPassword(ctx, cl, "felis"); err != nil || pw != "" {
+			t.Errorf("smtpSecretPassword = (%q, %v), want (\"\", nil)", pw, err)
+		}
+	})
+
+	t.Run("any other read failure is an error", func(t *testing.T) {
+		cl := fake.NewClientBuilder().WithScheme(scheme).WithInterceptorFuncs(interceptor.Funcs{
+			Get: func(context.Context, client.WithWatch, client.ObjectKey, client.Object, ...client.GetOption) error {
+				return apierrors.NewForbidden(schema.GroupResource{Resource: "secrets"}, platform.SMTPSecretName, errors.New("rbac"))
+			},
+		}).Build()
+		if _, err := smtpSecretPassword(ctx, cl, "felis"); err == nil {
+			t.Fatal("want the read failure")
+		}
+	})
+}
+
+// recoveryModel is an Owner-reset console for admin "root" (verified address
+// [email protected]), run by OS user alice, with the fake relay behind it.
+func recoveryModel(t *testing.T, f *fakeOwnerStore, r *fakeRelay, now func() time.Time) *ownerModel {
+	t.Helper()
+	if f.users == nil {
+		f.users = map[string]*api.StaffUser{"root": verifiedAdmin("root", "[email protected]")}
+	}
+	return newOwnerModel(context.Background(), f, "alice", true).withRecovery(relayConfig(r, now))
+}
+
+// nameAdmin submits the admin-name form and feeds the result of the send back in.
+func nameAdmin(t *testing.T, m *ownerModel, name string) *ownerModel {
+	t.Helper()
+	m.authUser = name
+	_, cmd := m.onFormComplete()
+	if m.step != owWorking {
+		t.Fatalf("after naming the admin step = %v, want owWorking", m.step)
+	}
+	msg := findMsg[owAuthMsg](t, cmd)
+	next, _ := m.Update(msg)
+	return next.(*ownerModel)
+}
+
+// findMsg runs a (possibly batched) command and returns the first T it produces.
+func findMsg[T any](t *testing.T, cmd tea.Cmd) T {
+	t.Helper()
+	var zero T
+	if cmd == nil {
+		t.Fatalf("no command, want one producing %T", zero)
+	}
+	switch msg := cmd().(type) {
+	case T:
+		return msg
+	case tea.BatchMsg:
+		for _, c := range msg {
+			if c == nil {
+				continue
+			}
+			if got, ok := c().(T); ok {
+				return got
+			}
+		}
+	}
+	t.Fatalf("command produced no %T", zero)
+	return zero
+}
+
+// typeCode submits the code form with typed.
+func typeCode(t *testing.T, m *ownerModel, typed string) {
+	t.Helper()
+	if m.step != owCode {
+		t.Fatalf("step = %v, want owCode", m.step)
+	}
+	m.codeInput = typed
+	m.onFormComplete()
+}
+
+// provisionAudit finishes the run as an Owner reset and returns its audit payload.
+func provisionAudit(t *testing.T, m *ownerModel, f *fakeOwnerStore) (api.AuditEntry, map[string]any) {
+	t.Helper()
+	if m.step != owProvision {
+		t.Fatalf("step = %v, want owProvision", m.step)
+	}
+	m.username = "owner"
+	msg := m.provisionCmd()().(owProvisionMsg)
+	if msg.err != nil {
+		t.Fatalf("provision: %v", msg.err)
+	}
+	return auditOf(t, f)
+}
+
+func TestRecoveryConsoleFlow(t *testing.T) {
+	t0 := time.Date(2026, 9, 26, 8, 0, 0, 0, time.UTC)
+	fixed := func() time.Time { return t0 }
+
+	t.Run("the mailed code proves the admin and the audit says so", func(t *testing.T) {
+		f, r := &fakeOwnerStore{}, &fakeRelay{}
+		m := nameAdmin(t, recoveryModel(t, f, r, fixed), "root")
+		typeCode(t, m, r.sentCode(t))
+		if m.mode != "recovery" || m.accountable != "root" {
+			t.Fatalf("mode=%q accountable=%q, want recovery attributed to root", m.mode, m.accountable)
+		}
+		e, payload := provisionAudit(t, m, f)
+		if e.Actor != "root" || e.Action != "break_glass.recovery" {
+			t.Errorf("audit = %+v, want actor=root action=break_glass.recovery", e)
+		}
+		if payload["verified"] != true || payload["verified_by"] != verifiedByEmailOTP || payload["code_sent_to"] != "[email protected]" || payload["os_user"] != "alice" {
+			t.Errorf("payload = %v, want verified by email_otp to [email protected], os_user alice", payload)
+		}
+	})
+
+	t.Run("a wrong code asks again, and the last wrong one leads to the override", func(t *testing.T) {
+		f, r := &fakeOwnerStore{}, &fakeRelay{}
+		m := nameAdmin(t, recoveryModel(t, f, r, fixed), "root")
+		wrong := "000000"
+		if r.sentCode(t) == wrong {
+			wrong = "111111"
+		}
+		for i := 1; i < recoveryCodeAttempts; i++ {
+			typeCode(t, m, wrong)
+			if m.step != owCode || !strings.Contains(m.codeNote, "wrong") {
+				t.Fatalf("wrong code %d: step=%v note=%q, want the code form again with a note", i, m.step, m.codeNote)
+			}
+		}
+		typeCode(t, m, wrong)
+		if m.step != owOverride || m.skip != otpSkipCodeRejected {
+			t.Fatalf("after %d wrong codes step=%v skip=%q, want the override for code_rejected", recoveryCodeAttempts, m.step, m.skip)
+		}
+		m.onFormComplete() // OVERRIDE typed
+		e, payload := provisionAudit(t, m, f)
+		if e.Actor != "alice" || e.Action != "break_glass.root_override" {
+			t.Errorf("audit = %+v, want actor=alice action=break_glass.root_override", e)
+		}
+		if payload["verified"] != false || payload["otp_skipped"] != otpSkipCodeRejected || payload["admin_account"] != "root" {
+			t.Errorf("payload = %v, want unverified, otp_skipped=code_rejected, admin_account=root", payload)
+		}
+	})
+
+	t.Run("a late code leads to the override", func(t *testing.T) {
+		now := t0
+		f, r := &fakeOwnerStore{}, &fakeRelay{}
+		m := nameAdmin(t, recoveryModel(t, f, r, func() time.Time { return now }), "root")
+		now = t0.Add(recoveryCodeTTL)
+		typeCode(t, m, r.sentCode(t))
+		if m.step != owOverride || m.skip != otpSkipCodeExpired {
+			t.Fatalf("step=%v skip=%q, want the override for code_expired", m.step, m.skip)
+		}
+	})
+
+	t.Run("OVERRIDE at the code prompt goes on unverified, saying the operator skipped", func(t *testing.T) {
+		f, r := &fakeOwnerStore{}, &fakeRelay{}
+		m := nameAdmin(t, recoveryModel(t, f, r, fixed), "root")
+		typeCode(t, m, breakGlassOverrideToken)
+		if m.mode != "root_override" || m.accountable != "alice" {
+			t.Fatalf("mode=%q accountable=%q, want root_override as alice", m.mode, m.accountable)
+		}
+		_, payload := provisionAudit(t, m, f)
+		if payload["verified"] != false || payload["otp_skipped"] != otpSkipByOperator {
+			t.Errorf("payload = %v, want unverified, otp_skipped=operator_skipped", payload)
+		}
+	})
+
+	t.Run("a relay failure leads to the override naming it", func(t *testing.T) {
+		f, r := &fakeOwnerStore{}, &fakeRelay{err: errors.New("dial tcp 10.0.0.9:587: connect: connection refused")}
+		m := nameAdmin(t, recoveryModel(t, f, r, fixed), "root")
+		if m.step != owOverride || m.skip != otpSkipSendFailed {
+			t.Fatalf("step=%v skip=%q, want the override for send_failed", m.step, m.skip)
+		}
+		if reason := m.overrideReason(); !strings.Contains(reason, "connection refused") {
+			t.Errorf("override reason %q does not name the failure", reason)
+		}
+		m.onFormComplete()
+		_, payload := provisionAudit(t, m, f)
+		if payload["otp_skipped"] != otpSkipSendFailed || !strings.Contains(payload["otp_skip_detail"].(string), "connection refused") {
+			t.Errorf("payload = %v, want otp_skipped=send_failed with the relay's error", payload)
+		}
+	})
+
+	t.Run("esc at the code prompt starts over and forgets the code", func(t *testing.T) {
+		f, r := &fakeOwnerStore{}, &fakeRelay{}
+		m := nameAdmin(t, recoveryModel(t, f, r, fixed), "root")
+		code := r.sentCode(t)
+		next, _ := m.Update(key(tea.KeyEsc))
+		m = next.(*ownerModel)
+		if m.step != owAuth || m.code != nil || m.admin != nil {
+			t.Fatalf("after esc step=%v code=%v admin=%v, want owAuth with the attempt forgotten", m.step, m.code, m.admin)
+		}
+		// The old code cannot be replayed: the next name mails a new one.
+		m = nameAdmin(t, m, "root")
+		if len(r.sent) != 2 {
+			t.Fatalf("relay carried %d mails, want a second one for the new attempt", len(r.sent))
+		}
+		_, fresh, _ := strings.Cut(r.sent[1].body, "Recovery code: ")
+		if m.code.value != fresh[:6] {
+			t.Errorf("the console checks %q, want the newly mailed %q (old one was %q)", m.code.value, fresh[:6], code)
+		}
+	})
+}
+
+func TestRootHandsRecoveryToAccountOperations(t *testing.T) {
+	for _, op := range []bgOperation{bgProvisionOwner, bgAddOperator} {
+		m := newTestRoot(true, consoleModeBreakGlass, "")
+		m.recovery = recoveryConfig{host: "felis-host-1"}
+		m = drive(t, m, menuChoiceMsg{op: op})
+		om, ok := m.screen.(*ownerModel)
+		if !ok {
+			t.Fatalf("op %v: screen = %T, want *ownerModel", op, m.screen)
+		}
+		if om.recovery.host != "felis-host-1" {
+			t.Errorf("op %v: the account screen has no relay config; its codes could never go out", op)
+		}
+	}
+}

cmd/felis/breakglass_test.go

@@ -269,71 +269,57 @@ func TestEnableLocalAuth(t *testing.T) {
 	}
 }
 
-func TestAuthenticateAdmin(t *testing.T) {
+func TestResolveAdmin(t *testing.T) {
 	ctx := context.Background()
 
-	// Password verification is gone (passwordless design): authenticateAdmin now only
-	// resolves the named admin so recovery can attribute the audit to a real identity.
-	// The security boundary is the break-glass root gate, not a typed secret.
+	// resolveAdmin only finds the staff account a typed name points at; proving the
+	// operator holds it is the mailed code's job (beginRecovery).
 
-	t.Run("resolves an existing admin for attribution", func(t *testing.T) {
+	t.Run("resolves an existing admin", func(t *testing.T) {
 		f := &fakeOwnerStore{users: map[string]*api.StaffUser{"root": mkAdmin("root")}}
-		matched, ok, err := authenticateAdmin(ctx, f, "root")
+		u, err := resolveAdmin(ctx, f, "  root ")
 		if err != nil {
-			t.Fatalf("authenticateAdmin: %v", err)
+			t.Fatalf("resolveAdmin: %v", err)
 		}
-		if !ok {
-			t.Fatal("ok = false, want true for an existing admin")
-		}
-		if matched != "root" {
-			t.Errorf("matched = %q, want root", matched)
+		if u == nil || u.Username != "root" {
+			t.Fatalf("resolveAdmin = %+v, want the root admin", u)
 		}
 	})
 
-	t.Run("a non-admin role can never attribute a break-glass", func(t *testing.T) {
+	t.Run("a non-admin role is no staff account", func(t *testing.T) {
 		player := mkAdmin("alice")
 		player.Role = "user" // a player row is not staff
 		f := &fakeOwnerStore{users: map[string]*api.StaffUser{"alice": player}}
-		_, ok, err := authenticateAdmin(ctx, f, "alice")
-		if err != nil {
-			t.Fatalf("unexpected error: %v", err)
-		}
-		if ok {
-			t.Error("ok = true, want false for a non-admin role")
+		if u, err := resolveAdmin(ctx, f, "alice"); u != nil || err != nil {
+			t.Errorf("resolveAdmin(player) = (%+v, %v), want (nil, nil)", u, err)
 		}
 	})
 
-	t.Run("the owner role attributes like an admin", func(t *testing.T) {
+	t.Run("the owner role counts as staff", func(t *testing.T) {
 		owner := mkAdmin("root")
 		owner.Role = "owner" // the platform owner is staff too (migration 0011)
 		f := &fakeOwnerStore{users: map[string]*api.StaffUser{"root": owner}}
-		matched, ok, err := authenticateAdmin(ctx, f, "root")
-		if err != nil || !ok || matched != "root" {
-			t.Fatalf("authenticateAdmin(owner) = (%q, %v, %v), want (root, true, nil)", matched, ok, err)
+		if u, err := resolveAdmin(ctx, f, "root"); err != nil || u == nil {
+			t.Fatalf("resolveAdmin(owner) = (%+v, %v), want the owner", u, err)
 		}
 	})
 
-	t.Run("an unknown user is a non-match, not an error", func(t *testing.T) {
-		f := &fakeOwnerStore{}
-		_, ok, err := authenticateAdmin(ctx, f, "nobody")
-		if err != nil {
-			t.Fatalf("unexpected error: %v", err)
-		}
-		if ok {
-			t.Error("ok = true, want false for an unknown user")
+	t.Run("an unknown user is nil, not an error", func(t *testing.T) {
+		if u, err := resolveAdmin(ctx, &fakeOwnerStore{}, "nobody"); u != nil || err != nil {
+			t.Errorf("resolveAdmin(unknown) = (%+v, %v), want (nil, nil)", u, err)
 		}
 	})
 
-	t.Run("an empty username is a non-match with no store call", func(t *testing.T) {
+	t.Run("an empty username makes no store call", func(t *testing.T) {
 		f := &fakeOwnerStore{userErr: errors.New("must not be called")}
-		if _, ok, err := authenticateAdmin(ctx, f, ""); ok || err != nil {
-			t.Errorf("empty username: ok=%v err=%v, want false,nil", ok, err)
+		if u, err := resolveAdmin(ctx, f, " "); u != nil || err != nil {
+			t.Errorf("empty username: (%+v, %v), want (nil, nil)", u, err)
 		}
 	})
 
 	t.Run("a datastore fault is surfaced", func(t *testing.T) {
 		f := &fakeOwnerStore{userErr: errors.New("db down")}
-		if _, _, err := authenticateAdmin(ctx, f, "root"); err == nil {
+		if _, err := resolveAdmin(ctx, f, "root"); err == nil {
 			t.Fatal("want error when the store fails")
 		}
 	})
@@ -401,6 +387,8 @@ func TestPerformBreakGlass(t *testing.T) {
 			osUser:         "alice",
 			ownerUsername:  "owner",
 			attemptedAdmin: "root",
+			verifiedBy:     verifiedByEmailOTP,
+			codeSentTo:     "[email protected]",
 		}
 		out, err := performBreakGlass(ctx, f, op)
 		if err != nil {
@@ -425,6 +413,23 @@ func TestPerformBreakGlass(t *testing.T) {
 		if payload["admin_account"] != "root" {
 			t.Errorf("payload.admin_account = %v, want root", payload["admin_account"])
 		}
+		if payload["verified_by"] != verifiedByEmailOTP || payload["code_sent_to"] != "[email protected]" {
+			t.Errorf("payload = %v, want verified_by=email_otp [email protected]", payload)
+		}
+		if _, present := payload["otp_skipped"]; present {
+			t.Error("a proven recovery carries no otp_skipped")
+		}
+	})
+
+	t.Run("recovery without a proof is recorded unverified", func(t *testing.T) {
+		f := &fakeOwnerStore{}
+		op := breakGlassOp{mode: "recovery", accountable: "root", osUser: "alice", ownerUsername: "owner", attemptedAdmin: "root"}
+		if _, err := performBreakGlass(ctx, f, op); err != nil {
+			t.Fatalf("performBreakGlass: %v", err)
+		}
+		if _, payload := auditOf(t, f); payload["verified"] != false {
+			t.Errorf("payload.verified = %v, want false: only a mailed code verifies", payload["verified"])
+		}
 	})
 
 	t.Run("root override records an unverified row attributed to the OS user", func(t *testing.T) {
@@ -435,6 +440,8 @@ func TestPerformBreakGlass(t *testing.T) {
 			osUser:         "alice",
 			ownerUsername:  "owner",
 			attemptedAdmin: "typo-admin",
+			otpSkipped:     otpSkipSendFailed,
+			otpSkipDetail:  "dial tcp: connection refused",
 		}
 		if _, err := performBreakGlass(ctx, f, op); err != nil {
 			t.Fatalf("performBreakGlass: %v", err)
@@ -450,6 +457,13 @@ func TestPerformBreakGlass(t *testing.T) {
 		if payload["admin_account"] != "typo-admin" {
 			t.Errorf("payload.admin_account = %v, want typo-admin", payload["admin_account"])
 		}
+		// Why no code proved anyone is part of the record.
+		if payload["otp_skipped"] != otpSkipSendFailed || payload["otp_skip_detail"] != "dial tcp: connection refused" {
+			t.Errorf("payload = %v, want otp_skipped=send_failed with its detail", payload)
+		}
+		if _, present := payload["verified_by"]; present {
+			t.Error("an override carries no verified_by")
+		}
 	})
 
 	t.Run("an audit failure does not fail the recovery", func(t *testing.T) {
@@ -616,6 +630,8 @@ func TestPerformAddOperator(t *testing.T) {
 			ownerUsername:  "ops-jordan",
 			ownerEmail:     "[email protected]",
 			attemptedAdmin: "root",
+			verifiedBy:     verifiedByEmailOTP,
+			codeSentTo:     "[email protected]",
 		}
 		out, err := performAddOperator(ctx, f, op)
 		if err != nil {
@@ -642,14 +658,14 @@ func TestPerformAddOperator(t *testing.T) {
 		if _, present := payload["owner"]; present {
 			t.Error("payload.owner present, want the new account under the operator key")
 		}
-		if payload["verified"] != true || payload["admin_account"] != "root" {
-			t.Errorf("payload = %v, want verified=true admin_account=root", payload)
+		if payload["verified"] != true || payload["admin_account"] != "root" || payload["verified_by"] != verifiedByEmailOTP {
+			t.Errorf("payload = %v, want verified=true admin_account=root verified_by=email_otp", payload)
 		}
 	})
 
 	t.Run("root override records an unverified operator row", func(t *testing.T) {
 		f := &fakeOwnerStore{}
-		op := breakGlassOp{mode: "root_override", accountable: "alice", osUser: "alice", ownerUsername: "ops", attemptedAdmin: "typo-admin"}
+		op := breakGlassOp{mode: "root_override", accountable: "alice", osUser: "alice", ownerUsername: "ops", attemptedAdmin: "typo-admin", otpSkipped: otpSkipUnknownAdmin}
 		if _, err := performAddOperator(ctx, f, op); err != nil {
 			t.Fatalf("performAddOperator: %v", err)
 		}
@@ -657,8 +673,8 @@ func TestPerformAddOperator(t *testing.T) {
 			t.Fatalf("want 1 insert, got %d", len(f.inserts))
 		}
 		_, payload := auditOf(t, f)
-		if payload["verified"] != false {
-			t.Errorf("payload.verified = %v, want false for root_override", payload["verified"])
+		if payload["verified"] != false || payload["otp_skipped"] != otpSkipUnknownAdmin {
+			t.Errorf("payload = %v, want verified=false otp_skipped=unknown_admin", payload)
 		}
 	})
 

cmd/felis/tui_owner.go

@@ -5,6 +5,7 @@ import (
 	"errors"
 	"fmt"
 	"strings"
+	"time"
 
 	"felis.lolicon.best/internal/api"
 
@@ -14,9 +15,8 @@ import (
 )
 
 type owAuthMsg struct {
-	matched string
-	ok      bool
-	err     error
+	start recoveryStart
+	err   error
 }
 
 type owProvisionMsg struct {
@@ -29,14 +29,15 @@ type owStep int
 const (
 	owAuth owStep = iota
 	owOverride
+	owCode // typing the recovery code mailed to the named admin
 	owProvision
 	owWorking
 	owDone
 	owError
 )
 
-// ownerModel collects the owner account. The input phases (admin auth, root
-// override, owner details) are huh forms; the async phases (verifying,
+// ownerModel collects the owner account. The input phases (admin name, recovery
+// code, root override, owner details) are huh forms; the async phases (verifying,
 // provisioning) show a spinner; the done phase shows the credential card. The
 // outward contract is unchanged: it emits an ownerResultMsg when finished.
 //
@@ -55,6 +56,18 @@ type ownerModel struct {
 	mode        string // "bootstrap", "recovery", "root_override"
 	accountable string
 	attempt     string
+	recovery    recoveryConfig
+
+	// Recovery proof: the admin the typed name resolved to, the code mailed to it,
+	// and once settled either how it was proven or why the run fell back to the
+	// override.
+	admin      *api.StaffUser
+	code       *recoveryCode
+	codeNote   string // "wrong code" line shown above a rebuilt code form
+	verifiedBy string
+	codeSentTo string
+	skip       string // otpSkip*
+	skipDetail string
 
 	step         owStep
 	form         *huh.Form
@@ -66,6 +79,7 @@ type ownerModel struct {
 
 	// huh-bound form values
 	authUser    string
+	codeInput   string
 	overrideTok string
 	ownerUser   string
 	ownerEmail  string
@@ -124,6 +138,12 @@ func newOperatorModel(ctx context.Context, store ownerStore, osUser string) *own
 	return m
 }
 
+// withRecovery hands the model the relay its recovery codes go through.
+func (m *ownerModel) withRecovery(r recoveryConfig) *ownerModel {
+	m.recovery = r
+	return m
+}
+
 func (m *ownerModel) Init() tea.Cmd { return m.form.Init() }
 
 // subject is the human label for the account being provisioned, branching every
@@ -152,7 +172,7 @@ func (m *ownerModel) sized(f *huh.Form) *huh.Form {
 }
 
 func (m *ownerModel) isFormStep() bool {
-	return m.step == owAuth || m.step == owOverride || m.step == owProvision
+	return m.step == owAuth || m.step == owOverride || m.step == owCode || m.step == owProvision
 }
 
 func (m *ownerModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
@@ -161,16 +181,15 @@ func (m *ownerModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 		if msg.err != nil {
 			return m, m.failCmd(msg.err)
 		}
-		if msg.ok {
-			m.mode = "recovery"
-			m.accountable = msg.matched
-			m.step = owProvision
-			m.form = m.sized(m.buildProvisionForm())
+		m.admin = msg.start.admin
+		if msg.start.code != nil {
+			m.code = msg.start.code
+			m.codeNote = ""
+			m.step = owCode
+			m.form = m.sized(m.buildCodeForm())
 			return m, m.form.Init()
 		}
-		m.step = owOverride
-		m.form = m.sized(m.buildOverrideForm())
-		return m, m.form.Init()
+		return m.toOverride(msg.start.skip, msg.start.detail)
 
 	case owProvisionMsg:
 		if msg.err != nil {
@@ -221,7 +240,9 @@ func (m *ownerModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 			case "ctrl+c":
 				return m, tea.Quit
 			case "esc":
-				if m.step == owOverride {
+				if m.step == owOverride || m.step == owCode {
+					// Start over: a code mailed for the old attempt dies with it.
+					m.admin, m.code, m.skip, m.skipDetail = nil, nil, "", ""
 					m.step = owAuth
 					m.form = m.sized(m.buildAuthForm())
 					return m, m.form.Init()
@@ -253,18 +274,41 @@ func (m *ownerModel) onFormComplete() (tea.Model, tea.Cmd) {
 	case owAuth:
 		m.attempt = strings.TrimSpace(m.authUser)
 		m.step = owWorking
-		m.working = "Verifying admin…"
-		user := m.authUser
+		m.working = "Sending a recovery code…"
+		user, rc, osUser, op := m.authUser, m.recovery, m.osUser, m.operation
 		return m, tea.Batch(m.sp.Tick, func() tea.Msg {
-			matched, ok, err := authenticateAdmin(m.ctx, m.store, user)
-			return owAuthMsg{matched: matched, ok: ok, err: err}
+			start, err := beginRecovery(m.ctx, m.store, rc, user, osUser, op)
+			return owAuthMsg{start: start, err: err}
 		})
+	case owCode:
+		typed := strings.TrimSpace(m.codeInput)
+		m.codeInput = ""
+		if typed == breakGlassOverrideToken {
+			m.skip, m.skipDetail = otpSkipByOperator, ""
+			m.code = nil
+			return m.proceedAsRoot()
+		}
+		switch m.code.check(typed, m.recovery.clock()) {
+		case codeAccepted:
+			m.mode = "recovery"
+			m.accountable = m.admin.Username
+			m.verifiedBy = verifiedByEmailOTP
+			m.codeSentTo = m.admin.Email
+			m.code = nil
+			m.step = owProvision
+			m.form = m.sized(m.buildProvisionForm())
+			return m, m.form.Init()
+		case codeWrong:
+			m.codeNote = fmt.Sprintf("That code is wrong. %d attempts left.", m.code.attemptsLeft())
+			m.form = m.sized(m.buildCodeForm())
+			return m, m.form.Init()
+		case codeExpired:
+			return m.toOverride(otpSkipCodeExpired, "")
+		default:
+			return m.toOverride(otpSkipCodeRejected, fmt.Sprintf("%d wrong codes", recoveryCodeAttempts))
+		}
 	case owOverride:
-		m.mode = "root_override"
-		m.accountable = m.osUser
-		m.step = owProvision
-		m.form = m.sized(m.buildProvisionForm())
-		return m, m.form.Init()
+		return m.proceedAsRoot()
 	case owProvision:
 		m.username = strings.TrimSpace(m.ownerUser)
 		m.step = owWorking
@@ -274,6 +318,24 @@ func (m *ownerModel) onFormComplete() (tea.Model, tea.Cmd) {
 	return m, nil
 }
 
+// toOverride records why no code proved an admin and asks for the typed OVERRIDE.
+func (m *ownerModel) toOverride(skip, detail string) (tea.Model, tea.Cmd) {
+	m.skip, m.skipDetail = skip, detail
+	m.code = nil
+	m.step = owOverride
+	m.form = m.sized(m.buildOverrideForm())
+	return m, m.form.Init()
+}
+
+// proceedAsRoot is the typed OVERRIDE: the run goes on as the OS user, unverified.
+func (m *ownerModel) proceedAsRoot() (tea.Model, tea.Cmd) {
+	m.mode = "root_override"
+	m.accountable = m.osUser
+	m.step = owProvision
+	m.form = m.sized(m.buildProvisionForm())
+	return m, m.form.Init()
+}
+
 func (m *ownerModel) provisionCmd() tea.Cmd {
 	op := breakGlassOp{
 		mode:           m.mode,
@@ -282,6 +344,10 @@ func (m *ownerModel) provisionCmd() tea.Cmd {
 		ownerUsername:  m.username,
 		ownerEmail:     m.ownerEmail,
 		attemptedAdmin: m.attempt,
+		verifiedBy:     m.verifiedBy,
+		codeSentTo:     m.codeSentTo,
+		otpSkipped:     m.skip,
+		otpSkipDetail:  m.skipDetail,
 	}
 	// performAddOperator and performBreakGlass share a signature; the operation
 	// discriminator selects which one runs. The operator path is insert-only and
@@ -320,7 +386,7 @@ func (m *ownerModel) buildAuthForm() *huh.Form {
 	return m.sized(newFelisForm(huh.NewGroup(
 		huh.NewNote().
 			Title("Admin authentication").
-			Description("A staff account already exists. Identify yourself to continue."),
+			Description("A staff account already exists. Name yours: a one-time code goes to its verified email address."),
 		huh.NewInput().
 			Title("Admin username").
 			Value(&m.authUser).
@@ -328,11 +394,64 @@ func (m *ownerModel) buildAuthForm() *huh.Form {
 	)))
 }
 
+func (m *ownerModel) buildCodeForm() *huh.Form {
+	desc := fmt.Sprintf("A recovery code went to %s, the verified address of %q. It works for %d minutes.\n\n"+
+		"No mail? Type %s to go on as OS user %q with root authority; the audit log records that as an unverified override. Esc starts over.",
+		maskEmail(m.admin.Email), m.admin.Username, int(recoveryCodeTTL/time.Minute), breakGlassOverrideToken, m.osUser)
+	if m.codeNote != "" {
+		desc = m.codeNote + "\n\n" + desc
+	}
+	return m.sized(newFelisForm(huh.NewGroup(
+		huh.NewNote().Title("Email verification").Description(desc),
+		huh.NewInput().
+			Title("Recovery code").
+			Value(&m.codeInput).
+			Validate(func(s string) error {
+				s = strings.TrimSpace(s)
+				if s == breakGlassOverrideToken || isRecoveryCodeShape(s) {
+					return nil
+				}
+				return errors.New("enter the 6-digit code, or " + breakGlassOverrideToken)
+			}),
+	)))
+}
+
+func isRecoveryCodeShape(s string) bool {
+	if len(s) != 6 {
+		return false
+	}
+	for _, c := range s {
+		if c < '0' || c > '9' {
+			return false
+		}
+	}
+	return true
+}
+
+// overrideReason says why no code proved an admin, first line of the override form.
+func (m *ownerModel) overrideReason() string {
+	switch m.skip {
+	case otpSkipUnknownAdmin:
+		return fmt.Sprintf("No staff account is named %q.", m.attempt)
+	case otpSkipNoVerifiedEmail:
+		return fmt.Sprintf("%q has no verified email address, so no recovery code can reach it.", m.admin.Username)
+	case otpSkipNoRelay:
+		return "No mail relay can send a recovery code: " + m.skipDetail + "."
+	case otpSkipSendFailed:
+		return "The recovery code could not be sent: " + m.skipDetail + "."
+	case otpSkipCodeExpired:
+		return "The recovery code expired."
+	case otpSkipCodeRejected:
+		return fmt.Sprintf("%d wrong codes; that code no longer works.", recoveryCodeAttempts)
+	}
+	return "No admin was verified."
+}
+
 func (m *ownerModel) buildOverrideForm() *huh.Form {
 	return m.sized(newFelisForm(huh.NewGroup(
 		huh.NewNote().
 			Title("Root override").
-			Description(fmt.Sprintf("That credential did not match. Proceed as OS user %q with root authority by typing the confirmation token.", m.osUser)),
+			Description(m.overrideReason()+"\n\n"+fmt.Sprintf("Proceed as OS user %q with root authority by typing the confirmation token. The audit log records this run as an unverified root override and the reason above. Esc starts over.", m.osUser)),
 		huh.NewInput().
 			Title("Type "+breakGlassOverrideToken+" to confirm").
 			Value(&m.overrideTok).
@@ -352,14 +471,14 @@ func (m *ownerModel) buildProvisionForm() *huh.Form {
 	desc := fmt.Sprintf("Create the first Owner — recorded as OS user %q.", m.osUser)
 	switch m.mode {
 	case "recovery":
-		desc = fmt.Sprintf("Authenticated as %q.", m.accountable)
+		desc = fmt.Sprintf("Verified as %q by an email code.", m.accountable)
 	case "root_override":
 		desc = "Root override — the Owner will be reset."
 	}
 	if m.operation == bgAddOperator {
 		switch m.mode {
 		case "recovery":
-			desc = fmt.Sprintf("Add an Operator — authenticated as %q.", m.accountable)
+			desc = fmt.Sprintf("Add an Operator — verified as %q by an email code.", m.accountable)
 		case "root_override":
 			desc = "Add an Operator (root override)."
 		}

cmd/felis/tui_root.go

@@ -142,6 +142,7 @@ type rootModel struct {
 	accessAud   string
 	namespace   string // minecraft workload namespace (cfg.K8s.Namespace); target of the halt op
 	adminExists bool
+	recovery    recoveryConfig // how the account operations mail a recovery code
 }
 
 func newRootModel(ctx context.Context, store ownerStore, dbURL, rootDomain, adminHostname, panelHostname, accessAud, namespace, osUser string, adminExists bool, mode consoleMode) *rootModel {
@@ -221,7 +222,7 @@ func (m *rootModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 		m.stage = stageOwner
 		switch msg.op {
 		case bgAddOperator:
-			return m.adopt(newOperatorModel(m.ctx, m.store, m.osUser))
+			return m.adopt(newOperatorModel(m.ctx, m.store, m.osUser).withRecovery(m.recovery))
 		case bgHaltServer:
 			return m.adopt(newHaltModel(m.ctx, m.store, m.namespace, m.osUser))
 		case bgSyncBackup:
@@ -229,7 +230,7 @@ func (m *rootModel) Update(msg tea.Msg) (tea.Model, tea.Cmd) {
 			// Service + service token the peer dials live in the control namespace.
 			return m.adopt(newBackupModel(m.ctx, m.namespace, platform.DefaultControlNamespace, m.osUser))
 		default:
-			return m.adopt(newOwnerModel(m.ctx, m.store, m.osUser, m.adminExists))
+			return m.adopt(newOwnerModel(m.ctx, m.store, m.osUser, m.adminExists).withRecovery(m.recovery))
 		}
 
 	case haltResultMsg:

cmd/felis/watchdog.go

@@ -187,16 +187,26 @@ func usesMirroredScanDB(cfg *config.Config) bool {
 // forgets it when the Secret is gone (a relay without AUTH). An env var named by
 // [smtp] password_ref, when set, wins at send time instead.
 func refreshSMTPPassword(ctx context.Context, cl client.Client, ns string, state *watchdog.State, stderr io.Writer) {
+	password, err := smtpSecretPassword(ctx, cl, ns)
+	if err != nil {
+		fmt.Fprintf(stderr, "felis watchdog: read %s/%s (keeping the cached relay password): %v\n", ns, platform.SMTPSecretName, err)
+		return
+	}
+	state.SMTPPassword = password
+}
+
+// smtpSecretPassword reads the relay password from the felis-smtp Secret. A missing
+// Secret is a relay without AUTH and reads as "".
+func smtpSecretPassword(ctx context.Context, cl client.Client, ns string) (string, error) {
 	var sec corev1.Secret
 	err := cl.Get(ctx, client.ObjectKey{Namespace: ns, Name: platform.SMTPSecretName}, &sec)
-	switch {
-	case apierrors.IsNotFound(err):
-		state.SMTPPassword = ""
-	case err != nil:
-		fmt.Fprintf(stderr, "felis watchdog: read %s/%s (keeping the cached relay password): %v\n", ns, platform.SMTPSecretName, err)
-	default:
-		state.SMTPPassword = string(sec.Data[platform.SMTPSecretPasswordKey])
+	if apierrors.IsNotFound(err) {
+		return "", nil
 	}
+	if err != nil {
+		return "", err
+	}
+	return string(sec.Data[platform.SMTPSecretPasswordKey]), nil
 }
 
 // ownerEmails pings PostgreSQL and returns the verified addresses of the