修复误判死

backend/internal/service/tokens.go

@@ -966,15 +966,27 @@ func (s *TokenService) RefreshGrokLiveness(ctx context.Context) {
 		if it.Pool != "grok" || it.Dead || it.Status == "disabled" || strings.TrimSpace(it.Value) == "" {
 			continue
 		}
-		sub, serr := s.grok.FetchSubscription(ctx, it.Value)
+		// A grok sso can momentarily 401 (upstream blip / proxy hiccup / anti-bot)
+		// while still being fully valid, so a single auth failure must never kill a
+		// live account. Retry the subscription probe up to 3 times on 401/403; only
+		// if all 3 attempts still return ErrAuth do we treat the session as dead.
+		var sub *grok.Subscription
+		var serr error
+		for attempt := 1; attempt <= 3; attempt++ {
+			sub, serr = s.grok.FetchSubscription(ctx, it.Value)
+			if serr == nil || !errors.Is(serr, grok.ErrAuth) {
+				break
+			}
+			if attempt < 3 {
+				time.Sleep(2 * time.Second)
+			}
+		}
 		if serr != nil {
-			// Any probe error — including 401/403 (ErrAuth) — is treated as
+			if errors.Is(serr, grok.ErrAuth) {
-			// transient: leave the account as-is and retry next tick. A grok sso
+				// 3 consecutive 401/403 → the sso session is genuinely dead.
-			// can momentarily 401 (upstream blip / proxy hiccup / anti-bot) while
+				_, _ = s.tokens.Update(ctx, "grok", it.ID, map[string]any{"status": "disabled", "dead": true})
-			// still being fully valid, so a single auth failure must NOT kill a
+			}
-			// live account. Genuine membership loss is still caught below via an
+			continue // other transient upstream error → leave as-is, retry next tick
-			// INACTIVE/empty subscription (!sub.Member).
-			continue
 		}
 		if sub == nil || !sub.Member {
 			// no ACTIVE subscription (INACTIVE / empty) → membership lapsed → dead.