feat: publish v1.0.0-rc.36-custom.2

Fixed-group tokens with a backup group now take precedence over the session affinity skip-retry-on-failure rule, so retryable upstream errors (503) fall through to the backup group instead of stopping on the primary channel. Non-retryable errors, explicit channel pins, and auto groups are unchanged.

Adds controller regression tests for the affinity override and a build-time Go test gate.

service/channel_affinity.go

@@ -11,6 +11,7 @@ import (
 	"time"
 
 	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
 	"github.com/QuantumNous/new-api/model"
 	"github.com/QuantumNous/new-api/pkg/cachex"
 	"github.com/QuantumNous/new-api/relaykit/dto"
@@ -627,6 +628,17 @@ func ShouldSkipRetryAfterChannelAffinityFailure(c *gin.Context) bool {
 	if c == nil {
 		return false
 	}
+	// Explicit fixed-token fallback takes precedence over cache affinity.
+	primary := common.GetContextKeyString(c, constant.ContextKeyTokenGroup)
+	if primary != "" && primary != "auto" && common.GetContextKeyBool(c, constant.ContextKeyTokenCrossGroupRetry) {
+		if groups, ok := common.GetContextKeyType[[]string](c, constant.ContextKeyTokenAutoGroups); ok {
+			for _, group := range FilterUserTokenAutoGroups(common.GetContextKeyString(c, constant.ContextKeyUserGroup), groups) {
+				if group != primary {
+					return false
+				}
+			}
+		}
+	}
 	v, ok := c.Get(ginKeyChannelAffinitySkipRetry)
 	if ok {
 		b, ok := v.(bool)

service/channel_select_auto_groups_test.go

@@ -193,3 +193,35 @@ func TestCacheGetRandomSatisfiedChannelSkipsUnavailableFixedPrimary(t *testing.T
 	assert.Equal(t, "vip", selectedGroup)
 	assert.Equal(t, "vip", common.GetContextKeyString(ctx, constant.ContextKeyAutoGroup))
 }
+
+func TestFixedBackupAfterAffinityFailureWithOneRetry(t *testing.T) {
+	db := setupChannelSelectAutoGroupsTest(t)
+	common.RetryTimes = 1
+	const modelName = "fixed-backup-affinity"
+	createChannelSelectAutoGroupsChannel(t, db, 13, "default", modelName)
+	createChannelSelectAutoGroupsChannel(t, db, 14, "vip", modelName)
+	model.InitChannelCache()
+	ctx, _ := gin.CreateTestContext(httptest.NewRecorder())
+	common.SetContextKey(ctx, constant.ContextKeyUserGroup, "default")
+	common.SetContextKey(ctx, constant.ContextKeyTokenGroup, "default")
+	common.SetContextKey(ctx, constant.ContextKeyTokenAutoGroups, []string{"vip"})
+	common.SetContextKey(ctx, constant.ContextKeyTokenCrossGroupRetry, true)
+	ctx.Set(ginKeyChannelAffinitySkipRetry, true)
+	require.False(t, ShouldSkipRetryAfterChannelAffinityFailure(ctx))
+
+	// The distributor's affinity hit bypasses the ordered group selector.
+	param := &RetryParam{Ctx: ctx, TokenGroup: "default", ModelName: modelName}
+	param.SetRetry(1)
+	first, _, err := CacheGetRandomSatisfiedChannel(param)
+	require.NoError(t, err)
+	require.NotNil(t, first)
+	require.Equal(t, 13, first.Id)
+	require.Equal(t, 0, param.GetRetry())
+	param.IncreaseRetry()
+	backup, group, err := CacheGetRandomSatisfiedChannel(param)
+	require.NoError(t, err)
+	require.NotNil(t, backup)
+	require.Equal(t, 14, backup.Id)
+	require.Equal(t, "vip", group)
+	require.Equal(t, "vip", common.GetContextKeyString(ctx, constant.ContextKeyAutoGroup))
+}