diff --git a/backend/internal/service/openai_live.go b/backend/internal/service/openai_live.go
index 3687c0c44..413d288aa 100644
--- a/backend/internal/service/openai_live.go
+++ b/backend/internal/service/openai_live.go
@@ -554,18 +554,26 @@ func (s *OpenAIGatewayService) ProxyLiveSideband(
 	runErr := s.runLiveController(proxyCtx, record, upstream, errCh)
 	cancel()
 	_, _ = store.ReleaseLiveController(context.Background(), record.CallHash, owner)
-	if errors.Is(runErr, ErrLiveCallNotFound) {
+	if liveSessionEnded(runErr) || !time.Now().Before(record.ExpiresAt) {
 		s.finalizeLiveCall(record)
 		return runErr
 	}
-	if !errors.Is(runErr, context.DeadlineExceeded) && time.Now().Before(record.ExpiresAt) {
-		go s.observeLiveCall(record.CallHash)
-		return runErr
-	}
-	s.finalizeLiveCall(record)
+	go s.observeLiveCall(record.CallHash)
 	return runErr
 }
 
+// liveSessionEnded 判断控制连接的退出原因是否意味着会话已终结（应 finalize：写
+// usage log 并释放租约），而不是可以交给 observer 重连的临时错误。
+//
+// ErrLiveUnavailable 在控制循环里只会来自租约续租失败。RefreshLiveLease 的 Lua 在
+// leaseID 被 GC 后不会重新写入，重连也拿不回并发槽 —— 若按临时错误重试，会话会以
+// 约 1 秒一轮的节奏空转到 ExpiresAt，期间持着上游连接却不计入任何并发限制。
+func liveSessionEnded(err error) bool {
+	return errors.Is(err, ErrLiveCallNotFound) ||
+		errors.Is(err, ErrLiveUnavailable) ||
+		errors.Is(err, context.DeadlineExceeded)
+}
+
 func (s *OpenAIGatewayService) runLiveController(
 	ctx context.Context,
 	record *LiveCallRecord,
@@ -626,7 +634,7 @@ func (s *OpenAIGatewayService) observeLiveCall(callHash string) {
 		if errors.Is(runErr, ErrLiveControllerChanged) {
 			return
 		}
-		if errors.Is(runErr, context.DeadlineExceeded) || errors.Is(runErr, ErrLiveCallNotFound) {
+		if liveSessionEnded(runErr) {
 			s.finalizeLiveCall(record)
 			return
 		}
@@ -706,7 +714,9 @@ func (s *OpenAIGatewayService) waitForLiveObserverRetry(record *LiveCallRecord)
 		return false
 	}
 	controller, err := store.GetLiveController(context.Background(), record.CallHash)
-	return err == nil && controller == LiveControllerObserver && time.Now().Before(record.ExpiresAt)
+	// 过期不在此处判定：返回 true 让调用方回到循环顶部的过期分支，由它 finalize
+	// （写 usage log + 释放租约）。在这里直接返回 false 会让会话静默结束、不留记录。
+	return err == nil && controller == LiveControllerObserver
 }
 
 func (s *OpenAIGatewayService) refreshLiveLease(record *LiveCallRecord) bool {
diff --git a/backend/internal/service/openai_live_lifecycle_test.go b/backend/internal/service/openai_live_lifecycle_test.go
index 59571cf3c..7ad2c1b0f 100644
--- a/backend/internal/service/openai_live_lifecycle_test.go
+++ b/backend/internal/service/openai_live_lifecycle_test.go
@@ -4,6 +4,7 @@ import (
 	"context"
 	"encoding/json"
 	"errors"
+	"fmt"
 	"net/http"
 	"net/http/httptest"
 	"strings"
@@ -415,3 +416,54 @@ func TestProxyLiveSidebandForwardsTextAndBinary(t *testing.T) {
 	upstream.reads <- liveTestFrame{err: coderws.CloseError{Code: coderws.StatusNormalClosure}}
 	require.ErrorIs(t, <-proxyResult, ErrLiveCallNotFound)
 }
+
+// TestLiveSessionEndedTreatsLeaseLossAsTerminal 锁定：租约续租失败（ErrLiveUnavailable）
+// 必须判为会话终结。RefreshLiveLease 的 Lua 在 leaseID 被 GC 后不会重新写入，若把它
+// 当临时错误交给 observer 重连，会话会空转到 ExpiresAt 且不计入任何并发限制。
+func TestLiveSessionEndedTreatsLeaseLossAsTerminal(t *testing.T) {
+	cases := []struct {
+		name string
+		err  error
+		want bool
+	}{
+		{"租约丢失", ErrLiveUnavailable, true},
+		{"租约丢失（被包装）", fmt.Errorf("refresh live lease: %w", ErrLiveUnavailable), true},
+		{"上游报告会话已关闭", ErrLiveCallNotFound, true},
+		{"到达会话时长上限", context.DeadlineExceeded, true},
+		{"控制权被他人接管", ErrLiveControllerChanged, false},
+		{"临时读错误", errors.New("unexpected EOF"), false},
+		{"无错误", nil, false},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			require.Equal(t, tc.want, liveSessionEnded(tc.err))
+		})
+	}
+}
+
+// TestWaitForLiveObserverRetryLeavesExpiryToLoopFinalize 锁定：已过期但控制权仍在
+// observer 手上时返回 true，让调用方回到 observeLiveCall 循环顶部的过期分支去
+// finalize（写 usage log + 释放租约）。在此处直接返回 false 会让会话静默结束、不留记录。
+func TestWaitForLiveObserverRetryLeavesExpiryToLoopFinalize(t *testing.T) {
+	record := &LiveCallRecord{
+		CallID:     "call_expired",
+		CallHash:   hashLiveCallID("call_expired"),
+		Controller: LiveControllerObserver,
+		ExpiresAt:  time.Now().Add(-time.Minute),
+	}
+	store := &liveTestStore{}
+	require.NoError(t, store.SaveLiveCall(context.Background(), record, time.Hour))
+	svc := &OpenAIGatewayService{cache: store}
+
+	require.True(t, svc.waitForLiveObserverRetry(record),
+		"过期判定必须留给循环顶部，否则不会写 usage log")
+
+	// 控制权已被他人接管时仍必须停止重试，避免与新控制者抢同一个 call。
+	require.NoError(t, store.SaveLiveCall(context.Background(), &LiveCallRecord{
+		CallID:     record.CallID,
+		CallHash:   record.CallHash,
+		Controller: LiveControllerProxy,
+		ExpiresAt:  time.Now().Add(time.Hour),
+	}, time.Hour))
+	require.False(t, svc.waitForLiveObserverRetry(record))
+}
diff --git a/backend/migrations/186_allow_live_usage_request_type.sql b/backend/migrations/188_allow_live_usage_request_type.sql
similarity index 100%
rename from backend/migrations/186_allow_live_usage_request_type.sql
rename to backend/migrations/188_allow_live_usage_request_type.sql
diff --git a/backend/migrations/187_add_group_allow_live.sql b/backend/migrations/189_add_group_allow_live.sql
similarity index 100%
rename from backend/migrations/187_add_group_allow_live.sql
rename to backend/migrations/189_add_group_allow_live.sql