feat: smart retry max 1 attempt + clear sticky session on failure

- Change antigravitySmartRetryMaxAttempts from 3 to 1 to prevent repeated rate limiting and long waits - Clear sticky session binding (DeleteSessionAccountID) after smart retry exhaustion, so subsequent requests don't hit the same rate-limited account - Add flow diagrams to Forward/ForwardGemini doc comments - Add comprehensive unit tests covering: - Sticky session cleared on retry failure (429, 503, network error) - Sticky session NOT cleared on retry success - Sticky session NOT cleared for non-sticky requests (empty hash) - Sticky session NOT cleared on long delay path (handled by handler) - Nil cache safety (no panic) - MaxAttempts constant verification - End-to-end retryLoop → switchError propagation with session clear
2026-02-07 19:16:59 +08:00
parent e3748da860
commit 3077fd279d
2 changed files with 678 additions and 30 deletions
--- a/backend/internal/service/antigravity_gateway_service.go
+++ b/backend/internal/service/antigravity_gateway_service.go
@@ -35,7 +35,7 @@ const (
 	// - 预检查：剩余限流时间 < 此阈值时等待，>= 此阈值时切换账号
 	antigravityRateLimitThreshold       = 7 * time.Second
 	antigravitySmartRetryMinWait        = 1 * time.Second  // 智能重试最小等待时间
-	antigravitySmartRetryMaxAttempts    = 3                // 智能重试最大次数
+	antigravitySmartRetryMaxAttempts    = 1                // 智能重试最大次数（仅重试 1 次，防止重复限流/长期等待）
 	antigravityDefaultRateLimitDuration = 30 * time.Second // 默认限流时间（无 retryDelay 时使用）

 	// Google RPC 状态和类型常量
@@ -247,6 +247,11 @@ func (s *AntigravityGatewayService) handleSmartRetry(p antigravityRetryLoopParam
 			}
 		}

+		// 清除粘性会话绑定，避免下次请求仍命中限流账号
+		if s.cache != nil && p.sessionHash != "" {
+			_ = s.cache.DeleteSessionAccountID(p.ctx, p.groupID, p.sessionHash)
+		}
+
 		// 返回账号切换信号，让上层切换账号重试
 		return &smartRetryResult{
 			action: smartRetryActionBreakWithResp,
@@ -952,6 +957,16 @@ func isModelNotFoundError(statusCode int, body []byte) bool {
 }

 // Forward 转发 Claude 协议请求（Claude → Gemini 转换）
+//
+// 限流处理流程:
+//
+//	请求 → antigravityRetryLoop → 预检查(remaining>0? → 切换账号) → 发送上游
+//	  ├─ 成功 → 正常返回
+//	  └─ 429/503 → handleSmartRetry
+//	      ├─ retryDelay >= 7s → 设置模型限流 + 清除粘性绑定 → 切换账号
+//	      └─ retryDelay <  7s → 等待后重试 1 次
+//	          ├─ 成功 → 正常返回
+//	          └─ 失败 → 设置模型限流 + 清除粘性绑定 → 切换账号
 func (s *AntigravityGatewayService) Forward(ctx context.Context, c *gin.Context, account *Account, body []byte, isStickySession bool) (*ForwardResult, error) {
 	startTime := time.Now()
 	sessionID := getSessionID(c)
@@ -1571,6 +1586,16 @@ func stripSignatureSensitiveBlocksFromClaudeRequest(req *antigravity.ClaudeReque
 }

 // ForwardGemini 转发 Gemini 协议请求
+//
+// 限流处理流程:
+//
+//	请求 → antigravityRetryLoop → 预检查(remaining>0? → 切换账号) → 发送上游
+//	  ├─ 成功 → 正常返回
+//	  └─ 429/503 → handleSmartRetry
+//	      ├─ retryDelay >= 7s → 设置模型限流 + 清除粘性绑定 → 切换账号
+//	      └─ retryDelay <  7s → 等待后重试 1 次
+//	          ├─ 成功 → 正常返回
+//	          └─ 失败 → 设置模型限流 + 清除粘性绑定 → 切换账号
 func (s *AntigravityGatewayService) ForwardGemini(ctx context.Context, c *gin.Context, account *Account, originalModel string, action string, stream bool, body []byte, isStickySession bool) (*ForwardResult, error) {
 	startTime := time.Now()
 	sessionID := getSessionID(c)
--- a/backend/internal/service/antigravity_smart_retry_test.go
+++ b/backend/internal/service/antigravity_smart_retry_test.go
@@ -13,6 +13,23 @@ import (
 	"github.com/stretchr/testify/require"
 )

+// stubSmartRetryCache 用于 handleSmartRetry 测试的 GatewayCache mock
+// 仅关注 DeleteSessionAccountID 的调用记录
+type stubSmartRetryCache struct {
+	GatewayCache // 嵌入接口，未实现的方法 panic（确保只调用预期方法）
+	deleteCalls  []deleteSessionCall
+}
+
+type deleteSessionCall struct {
+	groupID     int64
+	sessionHash string
+}
+
+func (c *stubSmartRetryCache) DeleteSessionAccountID(_ context.Context, groupID int64, sessionHash string) error {
+	c.deleteCalls = append(c.deleteCalls, deleteSessionCall{groupID: groupID, sessionHash: sessionHash})
+	return nil
+}
+
 // mockSmartRetryUpstream 用于 handleSmartRetry 测试的 mock upstream
 type mockSmartRetryUpstream struct {
 	responses []*http.Response
@@ -198,7 +215,7 @@ func TestHandleSmartRetry_ShortDelay_SmartRetrySuccess(t *testing.T) {

 // TestHandleSmartRetry_ShortDelay_SmartRetryFailed_ReturnsSwitchError 测试智能重试失败后返回 switchError
 func TestHandleSmartRetry_ShortDelay_SmartRetryFailed_ReturnsSwitchError(t *testing.T) {
-	// 智能重试后仍然返回 429（需要提供 3 个响应，因为智能重试最多 3 次）
+	// 智能重试后仍然返回 429（需要提供 1 个响应，因为智能重试最多 1 次）
 	failRespBody := `{
 		"error": {
 			"status": "RESOURCE_EXHAUSTED",
@@ -213,19 +230,9 @@ func TestHandleSmartRetry_ShortDelay_SmartRetryFailed_ReturnsSwitchError(t *test
 		Header:     http.Header{},
 		Body:       io.NopCloser(strings.NewReader(failRespBody)),
 	}
-	failResp2 := &http.Response{
-		StatusCode: http.StatusTooManyRequests,
-		Header:     http.Header{},
-		Body:       io.NopCloser(strings.NewReader(failRespBody)),
-	}
-	failResp3 := &http.Response{
-		StatusCode: http.StatusTooManyRequests,
-		Header:     http.Header{},
-		Body:       io.NopCloser(strings.NewReader(failRespBody)),
-	}
 	upstream := &mockSmartRetryUpstream{
-		responses: []*http.Response{failResp1, failResp2, failResp3},
-		errors:    []error{nil, nil, nil},
+		responses: []*http.Response{failResp1},
+		errors:    []error{nil},
 	}

 	repo := &stubAntigravityAccountRepo{}
@@ -236,7 +243,7 @@ func TestHandleSmartRetry_ShortDelay_SmartRetryFailed_ReturnsSwitchError(t *test
 		Platform: PlatformAntigravity,
 	}

-	// 3s < 7s 阈值，应该触发智能重试（最多 3 次）
+	// 3s < 7s 阈值，应该触发智能重试（最多 1 次）
 	respBody := []byte(`{
 		"error": {
 			"status": "RESOURCE_EXHAUSTED",
@@ -284,7 +291,7 @@ func TestHandleSmartRetry_ShortDelay_SmartRetryFailed_ReturnsSwitchError(t *test
 	// 验证模型限流已设置
 	require.Len(t, repo.modelRateLimitCalls, 1)
 	require.Equal(t, "gemini-3-flash", repo.modelRateLimitCalls[0].modelKey)
-	require.Len(t, upstream.calls, 3, "should have made three retry calls (max attempts)")
+	require.Len(t, upstream.calls, 1, "should have made one retry call (max attempts)")
 }

 // TestHandleSmartRetry_503_ModelCapacityExhausted_ReturnsSwitchError 测试 503 MODEL_CAPACITY_EXHAUSTED 返回 switchError
@@ -556,19 +563,15 @@ func TestAntigravityRetryLoop_HandleSmartRetry_SwitchError_Propagates(t *testing
 	require.True(t, switchErr.IsStickySession)
 }

-// TestHandleSmartRetry_NetworkError_ContinuesRetry 测试网络错误时继续重试
-func TestHandleSmartRetry_NetworkError_ContinuesRetry(t *testing.T) {
-	// 第一次网络错误，第二次成功
-	successResp := &http.Response{
-		StatusCode: http.StatusOK,
-		Header:     http.Header{},
-		Body:       io.NopCloser(strings.NewReader(`{"result":"ok"}`)),
-	}
+// TestHandleSmartRetry_NetworkError_ExhaustsRetry 测试网络错误时（maxAttempts=1）直接耗尽重试并切换账号
+func TestHandleSmartRetry_NetworkError_ExhaustsRetry(t *testing.T) {
+	// 唯一一次重试遇到网络错误（nil response）
 	upstream := &mockSmartRetryUpstream{
-		responses: []*http.Response{nil, successResp}, // 第一次返回 nil（模拟网络错误）
-		errors:    []error{nil, nil},                  // mock 不返回 error，靠 nil response 触发
+		responses: []*http.Response{nil}, // 返回 nil（模拟网络错误）
+		errors:    []error{nil},          // mock 不返回 error，靠 nil response 触发
 	}

+	repo := &stubAntigravityAccountRepo{}
 	account := &Account{
 		ID:       8,
 		Name:     "acc-8",
@@ -600,6 +603,7 @@ func TestHandleSmartRetry_NetworkError_ContinuesRetry(t *testing.T) {
 		action:       "generateContent",
 		body:         []byte(`{"input":"test"}`),
 		httpUpstream: upstream,
+		accountRepo:  repo,
 		handleError: func(ctx context.Context, prefix string, account *Account, statusCode int, headers http.Header, body []byte, quotaScope AntigravityQuotaScope, groupID int64, sessionHash string, isStickySession bool) *handleModelRateLimitResult {
 			return nil
 		},
@@ -612,10 +616,15 @@ func TestHandleSmartRetry_NetworkError_ContinuesRetry(t *testing.T) {

 	require.NotNil(t, result)
 	require.Equal(t, smartRetryActionBreakWithResp, result.action)
-	require.NotNil(t, result.resp, "should return successful response after network error recovery")
-	require.Equal(t, http.StatusOK, result.resp.StatusCode)
-	require.Nil(t, result.switchError, "should not return switchError on success")
-	require.Len(t, upstream.calls, 2, "should have made two retry calls")
+	require.Nil(t, result.resp, "should not return resp when switchError is set")
+	require.NotNil(t, result.switchError, "should return switchError after network error exhausted retry")
+	require.Equal(t, account.ID, result.switchError.OriginalAccountID)
+	require.Equal(t, "claude-sonnet-4-5", result.switchError.RateLimitedModel)
+	require.Len(t, upstream.calls, 1, "should have made one retry call")
+
+	// 验证模型限流已设置
+	require.Len(t, repo.modelRateLimitCalls, 1)
+	require.Equal(t, "claude-sonnet-4-5", repo.modelRateLimitCalls[0].modelKey)
 }

 // TestHandleSmartRetry_NoRetryDelay_UsesDefaultRateLimit 测试无 retryDelay 时使用默认 1 分钟限流
@@ -674,3 +683,617 @@ func TestHandleSmartRetry_NoRetryDelay_UsesDefaultRateLimit(t *testing.T) {
 	require.Len(t, repo.modelRateLimitCalls, 1)
 	require.Equal(t, "claude-sonnet-4-5", repo.modelRateLimitCalls[0].modelKey)
 }
+
+// ---------------------------------------------------------------------------
+// 以下测试覆盖本次改动：
+// 1. antigravitySmartRetryMaxAttempts = 1（仅重试 1 次）
+// 2. 智能重试失败后清除粘性会话绑定（DeleteSessionAccountID）
+// ---------------------------------------------------------------------------
+
+// TestSmartRetryMaxAttempts_VerifyConstant 验证常量值为 1
+func TestSmartRetryMaxAttempts_VerifyConstant(t *testing.T) {
+	require.Equal(t, 1, antigravitySmartRetryMaxAttempts,
+		"antigravitySmartRetryMaxAttempts should be 1 to prevent repeated rate limiting")
+}
+
+// TestHandleSmartRetry_ShortDelay_StickySession_FailedRetry_ClearsSession
+// 核心场景：粘性会话 + 短延迟重试失败 → 必须清除粘性绑定
+func TestHandleSmartRetry_ShortDelay_StickySession_FailedRetry_ClearsSession(t *testing.T) {
+	failRespBody := `{
+		"error": {
+			"status": "RESOURCE_EXHAUSTED",
+			"details": [
+				{"@type": "type.googleapis.com/google.rpc.ErrorInfo", "metadata": {"model": "claude-sonnet-4-5"}, "reason": "RATE_LIMIT_EXCEEDED"},
+				{"@type": "type.googleapis.com/google.rpc.RetryInfo", "retryDelay": "0.1s"}
+			]
+		}
+	}`
+	failResp := &http.Response{
+		StatusCode: http.StatusTooManyRequests,
+		Header:     http.Header{},
+		Body:       io.NopCloser(strings.NewReader(failRespBody)),
+	}
+	upstream := &mockSmartRetryUpstream{
+		responses: []*http.Response{failResp},
+		errors:    []error{nil},
+	}
+
+	repo := &stubAntigravityAccountRepo{}
+	cache := &stubSmartRetryCache{}
+	account := &Account{
+		ID:       10,
+		Name:     "acc-10",
+		Type:     AccountTypeOAuth,
+		Platform: PlatformAntigravity,
+	}
+
+	respBody := []byte(`{
+		"error": {
+			"status": "RESOURCE_EXHAUSTED",
+			"details": [
+				{"@type": "type.googleapis.com/google.rpc.ErrorInfo", "metadata": {"model": "claude-sonnet-4-5"}, "reason": "RATE_LIMIT_EXCEEDED"},
+				{"@type": "type.googleapis.com/google.rpc.RetryInfo", "retryDelay": "0.1s"}
+			]
+		}
+	}`)
+	resp := &http.Response{
+		StatusCode: http.StatusTooManyRequests,
+		Header:     http.Header{},
+		Body:       io.NopCloser(bytes.NewReader(respBody)),
+	}
+
+	params := antigravityRetryLoopParams{
+		ctx:             context.Background(),
+		prefix:          "[test]",
+		account:         account,
+		accessToken:     "token",
+		action:          "generateContent",
+		body:            []byte(`{"input":"test"}`),
+		httpUpstream:    upstream,
+		accountRepo:     repo,
+		isStickySession: true,
+		groupID:         42,
+		sessionHash:     "sticky-hash-abc",
+		handleError: func(ctx context.Context, prefix string, account *Account, statusCode int, headers http.Header, body []byte, quotaScope AntigravityQuotaScope, groupID int64, sessionHash string, isStickySession bool) *handleModelRateLimitResult {
+			return nil
+		},
+	}
+
+	availableURLs := []string{"https://ag-1.test"}
+
+	svc := &AntigravityGatewayService{cache: cache}
+	result := svc.handleSmartRetry(params, resp, respBody, "https://ag-1.test", 0, availableURLs)
+
+	// 验证返回 switchError
+	require.NotNil(t, result)
+	require.Equal(t, smartRetryActionBreakWithResp, result.action)
+	require.NotNil(t, result.switchError)
+	require.True(t, result.switchError.IsStickySession, "switchError should carry IsStickySession=true")
+	require.Equal(t, account.ID, result.switchError.OriginalAccountID)
+
+	// 核心断言：DeleteSessionAccountID 被调用，且参数正确
+	require.Len(t, cache.deleteCalls, 1, "should call DeleteSessionAccountID exactly once")
+	require.Equal(t, int64(42), cache.deleteCalls[0].groupID)
+	require.Equal(t, "sticky-hash-abc", cache.deleteCalls[0].sessionHash)
+
+	// 验证仅重试 1 次
+	require.Len(t, upstream.calls, 1, "should make exactly 1 retry call (maxAttempts=1)")
+
+	// 验证模型限流已设置
+	require.Len(t, repo.modelRateLimitCalls, 1)
+	require.Equal(t, "claude-sonnet-4-5", repo.modelRateLimitCalls[0].modelKey)
+}
+
+// TestHandleSmartRetry_ShortDelay_NonStickySession_FailedRetry_NoDeleteSession
+// 非粘性会话 + 短延迟重试失败 → 不应调用 DeleteSessionAccountID（sessionHash 为空）
+func TestHandleSmartRetry_ShortDelay_NonStickySession_FailedRetry_NoDeleteSession(t *testing.T) {
+	failRespBody := `{
+		"error": {
+			"status": "RESOURCE_EXHAUSTED",
+			"details": [
+				{"@type": "type.googleapis.com/google.rpc.ErrorInfo", "metadata": {"model": "gemini-3-flash"}, "reason": "RATE_LIMIT_EXCEEDED"},
+				{"@type": "type.googleapis.com/google.rpc.RetryInfo", "retryDelay": "0.1s"}
+			]
+		}
+	}`
+	failResp := &http.Response{
+		StatusCode: http.StatusTooManyRequests,
+		Header:     http.Header{},
+		Body:       io.NopCloser(strings.NewReader(failRespBody)),
+	}
+	upstream := &mockSmartRetryUpstream{
+		responses: []*http.Response{failResp},
+		errors:    []error{nil},
+	}
+
+	repo := &stubAntigravityAccountRepo{}
+	cache := &stubSmartRetryCache{}
+	account := &Account{
+		ID:       11,
+		Name:     "acc-11",
+		Type:     AccountTypeOAuth,
+		Platform: PlatformAntigravity,
+	}
+
+	respBody := []byte(`{
+		"error": {
+			"status": "RESOURCE_EXHAUSTED",
+			"details": [
+				{"@type": "type.googleapis.com/google.rpc.ErrorInfo", "metadata": {"model": "gemini-3-flash"}, "reason": "RATE_LIMIT_EXCEEDED"},
+				{"@type": "type.googleapis.com/google.rpc.RetryInfo", "retryDelay": "0.1s"}
+			]
+		}
+	}`)
+	resp := &http.Response{
+		StatusCode: http.StatusTooManyRequests,
+		Header:     http.Header{},
+		Body:       io.NopCloser(bytes.NewReader(respBody)),
+	}
+
+	params := antigravityRetryLoopParams{
+		ctx:             context.Background(),
+		prefix:          "[test]",
+		account:         account,
+		accessToken:     "token",
+		action:          "generateContent",
+		body:            []byte(`{"input":"test"}`),
+		httpUpstream:    upstream,
+		accountRepo:     repo,
+		isStickySession: false,
+		groupID:         42,
+		sessionHash:     "", // 非粘性会话，sessionHash 为空
+		handleError: func(ctx context.Context, prefix string, account *Account, statusCode int, headers http.Header, body []byte, quotaScope AntigravityQuotaScope, groupID int64, sessionHash string, isStickySession bool) *handleModelRateLimitResult {
+			return nil
+		},
+	}
+
+	availableURLs := []string{"https://ag-1.test"}
+
+	svc := &AntigravityGatewayService{cache: cache}
+	result := svc.handleSmartRetry(params, resp, respBody, "https://ag-1.test", 0, availableURLs)
+
+	require.NotNil(t, result)
+	require.Equal(t, smartRetryActionBreakWithResp, result.action)
+	require.NotNil(t, result.switchError)
+	require.False(t, result.switchError.IsStickySession)
+
+	// 核心断言：sessionHash 为空时不应调用 DeleteSessionAccountID
+	require.Len(t, cache.deleteCalls, 0, "should NOT call DeleteSessionAccountID when sessionHash is empty")
+}
+
+// TestHandleSmartRetry_ShortDelay_StickySession_FailedRetry_NilCache_NoPanic
+// 边界：cache 为 nil 时不应 panic
+func TestHandleSmartRetry_ShortDelay_StickySession_FailedRetry_NilCache_NoPanic(t *testing.T) {
+	failRespBody := `{
+		"error": {
+			"status": "RESOURCE_EXHAUSTED",
+			"details": [
+				{"@type": "type.googleapis.com/google.rpc.ErrorInfo", "metadata": {"model": "claude-sonnet-4-5"}, "reason": "RATE_LIMIT_EXCEEDED"},
+				{"@type": "type.googleapis.com/google.rpc.RetryInfo", "retryDelay": "0.1s"}
+			]
+		}
+	}`
+	failResp := &http.Response{
+		StatusCode: http.StatusTooManyRequests,
+		Header:     http.Header{},
+		Body:       io.NopCloser(strings.NewReader(failRespBody)),
+	}
+	upstream := &mockSmartRetryUpstream{
+		responses: []*http.Response{failResp},
+		errors:    []error{nil},
+	}
+
+	repo := &stubAntigravityAccountRepo{}
+	account := &Account{
+		ID:       12,
+		Name:     "acc-12",
+		Type:     AccountTypeOAuth,
+		Platform: PlatformAntigravity,
+	}
+
+	respBody := []byte(`{
+		"error": {
+			"status": "RESOURCE_EXHAUSTED",
+			"details": [
+				{"@type": "type.googleapis.com/google.rpc.ErrorInfo", "metadata": {"model": "claude-sonnet-4-5"}, "reason": "RATE_LIMIT_EXCEEDED"},
+				{"@type": "type.googleapis.com/google.rpc.RetryInfo", "retryDelay": "0.1s"}
+			]
+		}
+	}`)
+	resp := &http.Response{
+		StatusCode: http.StatusTooManyRequests,
+		Header:     http.Header{},
+		Body:       io.NopCloser(bytes.NewReader(respBody)),
+	}
+
+	params := antigravityRetryLoopParams{
+		ctx:             context.Background(),
+		prefix:          "[test]",
+		account:         account,
+		accessToken:     "token",
+		action:          "generateContent",
+		body:            []byte(`{"input":"test"}`),
+		httpUpstream:    upstream,
+		accountRepo:     repo,
+		isStickySession: true,
+		groupID:         42,
+		sessionHash:     "sticky-hash-nil-cache",
+		handleError: func(ctx context.Context, prefix string, account *Account, statusCode int, headers http.Header, body []byte, quotaScope AntigravityQuotaScope, groupID int64, sessionHash string, isStickySession bool) *handleModelRateLimitResult {
+			return nil
+		},
+	}
+
+	availableURLs := []string{"https://ag-1.test"}
+
+	// cache 为 nil，不应 panic
+	svc := &AntigravityGatewayService{cache: nil}
+	require.NotPanics(t, func() {
+		result := svc.handleSmartRetry(params, resp, respBody, "https://ag-1.test", 0, availableURLs)
+		require.NotNil(t, result)
+		require.Equal(t, smartRetryActionBreakWithResp, result.action)
+		require.NotNil(t, result.switchError)
+		require.True(t, result.switchError.IsStickySession)
+	})
+}
+
+// TestHandleSmartRetry_ShortDelay_StickySession_SuccessRetry_NoDeleteSession
+// 重试成功时不应清除粘性会话（只有失败才清除）
+func TestHandleSmartRetry_ShortDelay_StickySession_SuccessRetry_NoDeleteSession(t *testing.T) {
+	successResp := &http.Response{
+		StatusCode: http.StatusOK,
+		Header:     http.Header{},
+		Body:       io.NopCloser(strings.NewReader(`{"result":"ok"}`)),
+	}
+	upstream := &mockSmartRetryUpstream{
+		responses: []*http.Response{successResp},
+		errors:    []error{nil},
+	}
+
+	cache := &stubSmartRetryCache{}
+	account := &Account{
+		ID:       13,
+		Name:     "acc-13",
+		Type:     AccountTypeOAuth,
+		Platform: PlatformAntigravity,
+	}
+
+	respBody := []byte(`{
+		"error": {
+			"status": "RESOURCE_EXHAUSTED",
+			"details": [
+				{"@type": "type.googleapis.com/google.rpc.ErrorInfo", "metadata": {"model": "claude-opus-4"}, "reason": "RATE_LIMIT_EXCEEDED"},
+				{"@type": "type.googleapis.com/google.rpc.RetryInfo", "retryDelay": "0.5s"}
+			]
+		}
+	}`)
+	resp := &http.Response{
+		StatusCode: http.StatusTooManyRequests,
+		Header:     http.Header{},
+		Body:       io.NopCloser(bytes.NewReader(respBody)),
+	}
+
+	params := antigravityRetryLoopParams{
+		ctx:             context.Background(),
+		prefix:          "[test]",
+		account:         account,
+		accessToken:     "token",
+		action:          "generateContent",
+		body:            []byte(`{"input":"test"}`),
+		httpUpstream:    upstream,
+		isStickySession: true,
+		groupID:         42,
+		sessionHash:     "sticky-hash-success",
+		handleError: func(ctx context.Context, prefix string, account *Account, statusCode int, headers http.Header, body []byte, quotaScope AntigravityQuotaScope, groupID int64, sessionHash string, isStickySession bool) *handleModelRateLimitResult {
+			return nil
+		},
+	}
+
+	availableURLs := []string{"https://ag-1.test"}
+
+	svc := &AntigravityGatewayService{cache: cache}
+	result := svc.handleSmartRetry(params, resp, respBody, "https://ag-1.test", 0, availableURLs)
+
+	require.NotNil(t, result)
+	require.Equal(t, smartRetryActionBreakWithResp, result.action)
+	require.NotNil(t, result.resp, "should return successful response")
+	require.Equal(t, http.StatusOK, result.resp.StatusCode)
+	require.Nil(t, result.switchError, "should not return switchError on success")
+
+	// 核心断言：重试成功时不应清除粘性会话
+	require.Len(t, cache.deleteCalls, 0, "should NOT call DeleteSessionAccountID on successful retry")
+}
+
+// TestHandleSmartRetry_LongDelay_StickySession_NoDeleteInHandleSmartRetry
+// 长延迟路径（情况1）在 handleSmartRetry 中不直接调用 DeleteSessionAccountID
+// （清除由 handler 层的 shouldClearStickySession 在下次请求时处理）
+func TestHandleSmartRetry_LongDelay_StickySession_NoDeleteInHandleSmartRetry(t *testing.T) {
+	repo := &stubAntigravityAccountRepo{}
+	cache := &stubSmartRetryCache{}
+	account := &Account{
+		ID:       14,
+		Name:     "acc-14",
+		Type:     AccountTypeOAuth,
+		Platform: PlatformAntigravity,
+	}
+
+	// 15s >= 7s 阈值 → 走长延迟路径
+	respBody := []byte(`{
+		"error": {
+			"status": "RESOURCE_EXHAUSTED",
+			"details": [
+				{"@type": "type.googleapis.com/google.rpc.ErrorInfo", "metadata": {"model": "claude-sonnet-4-5"}, "reason": "RATE_LIMIT_EXCEEDED"},
+				{"@type": "type.googleapis.com/google.rpc.RetryInfo", "retryDelay": "15s"}
+			]
+		}
+	}`)
+	resp := &http.Response{
+		StatusCode: http.StatusTooManyRequests,
+		Header:     http.Header{},
+		Body:       io.NopCloser(bytes.NewReader(respBody)),
+	}
+
+	params := antigravityRetryLoopParams{
+		ctx:             context.Background(),
+		prefix:          "[test]",
+		account:         account,
+		accessToken:     "token",
+		action:          "generateContent",
+		body:            []byte(`{"input":"test"}`),
+		accountRepo:     repo,
+		isStickySession: true,
+		groupID:         42,
+		sessionHash:     "sticky-hash-long-delay",
+		handleError: func(ctx context.Context, prefix string, account *Account, statusCode int, headers http.Header, body []byte, quotaScope AntigravityQuotaScope, groupID int64, sessionHash string, isStickySession bool) *handleModelRateLimitResult {
+			return nil
+		},
+	}
+
+	availableURLs := []string{"https://ag-1.test"}
+
+	svc := &AntigravityGatewayService{cache: cache}
+	result := svc.handleSmartRetry(params, resp, respBody, "https://ag-1.test", 0, availableURLs)
+
+	require.NotNil(t, result)
+	require.Equal(t, smartRetryActionBreakWithResp, result.action)
+	require.NotNil(t, result.switchError)
+	require.True(t, result.switchError.IsStickySession)
+
+	// 长延迟路径不在 handleSmartRetry 中调用 DeleteSessionAccountID
+	// （由上游 handler 的 shouldClearStickySession 处理）
+	require.Len(t, cache.deleteCalls, 0,
+		"long delay path should NOT call DeleteSessionAccountID in handleSmartRetry (handled by handler layer)")
+}
+
+// TestHandleSmartRetry_ShortDelay_NetworkError_StickySession_ClearsSession
+// 网络错误耗尽重试 + 粘性会话 → 也应清除粘性绑定
+func TestHandleSmartRetry_ShortDelay_NetworkError_StickySession_ClearsSession(t *testing.T) {
+	upstream := &mockSmartRetryUpstream{
+		responses: []*http.Response{nil}, // 网络错误
+		errors:    []error{nil},
+	}
+
+	repo := &stubAntigravityAccountRepo{}
+	cache := &stubSmartRetryCache{}
+	account := &Account{
+		ID:       15,
+		Name:     "acc-15",
+		Type:     AccountTypeOAuth,
+		Platform: PlatformAntigravity,
+	}
+
+	respBody := []byte(`{
+		"error": {
+			"status": "RESOURCE_EXHAUSTED",
+			"details": [
+				{"@type": "type.googleapis.com/google.rpc.ErrorInfo", "metadata": {"model": "gemini-3-flash"}, "reason": "RATE_LIMIT_EXCEEDED"},
+				{"@type": "type.googleapis.com/google.rpc.RetryInfo", "retryDelay": "0.1s"}
+			]
+		}
+	}`)
+	resp := &http.Response{
+		StatusCode: http.StatusTooManyRequests,
+		Header:     http.Header{},
+		Body:       io.NopCloser(bytes.NewReader(respBody)),
+	}
+
+	params := antigravityRetryLoopParams{
+		ctx:             context.Background(),
+		prefix:          "[test]",
+		account:         account,
+		accessToken:     "token",
+		action:          "generateContent",
+		body:            []byte(`{"input":"test"}`),
+		httpUpstream:    upstream,
+		accountRepo:     repo,
+		isStickySession: true,
+		groupID:         99,
+		sessionHash:     "sticky-net-error",
+		handleError: func(ctx context.Context, prefix string, account *Account, statusCode int, headers http.Header, body []byte, quotaScope AntigravityQuotaScope, groupID int64, sessionHash string, isStickySession bool) *handleModelRateLimitResult {
+			return nil
+		},
+	}
+
+	availableURLs := []string{"https://ag-1.test"}
+
+	svc := &AntigravityGatewayService{cache: cache}
+	result := svc.handleSmartRetry(params, resp, respBody, "https://ag-1.test", 0, availableURLs)
+
+	require.NotNil(t, result)
+	require.NotNil(t, result.switchError)
+	require.True(t, result.switchError.IsStickySession)
+
+	// 核心断言：网络错误耗尽重试后也应清除粘性绑定
+	require.Len(t, cache.deleteCalls, 1, "should call DeleteSessionAccountID after network error exhausts retry")
+	require.Equal(t, int64(99), cache.deleteCalls[0].groupID)
+	require.Equal(t, "sticky-net-error", cache.deleteCalls[0].sessionHash)
+}
+
+// TestHandleSmartRetry_ShortDelay_503_StickySession_FailedRetry_ClearsSession
+// 503 + 短延迟 + 粘性会话 + 重试失败 → 清除粘性绑定
+func TestHandleSmartRetry_ShortDelay_503_StickySession_FailedRetry_ClearsSession(t *testing.T) {
+	failRespBody := `{
+		"error": {
+			"code": 503,
+			"status": "UNAVAILABLE",
+			"details": [
+				{"@type": "type.googleapis.com/google.rpc.ErrorInfo", "metadata": {"model": "gemini-3-pro"}, "reason": "MODEL_CAPACITY_EXHAUSTED"},
+				{"@type": "type.googleapis.com/google.rpc.RetryInfo", "retryDelay": "0.5s"}
+			]
+		}
+	}`
+	failResp := &http.Response{
+		StatusCode: http.StatusServiceUnavailable,
+		Header:     http.Header{},
+		Body:       io.NopCloser(strings.NewReader(failRespBody)),
+	}
+	upstream := &mockSmartRetryUpstream{
+		responses: []*http.Response{failResp},
+		errors:    []error{nil},
+	}
+
+	repo := &stubAntigravityAccountRepo{}
+	cache := &stubSmartRetryCache{}
+	account := &Account{
+		ID:       16,
+		Name:     "acc-16",
+		Type:     AccountTypeOAuth,
+		Platform: PlatformAntigravity,
+	}
+
+	respBody := []byte(`{
+		"error": {
+			"code": 503,
+			"status": "UNAVAILABLE",
+			"details": [
+				{"@type": "type.googleapis.com/google.rpc.ErrorInfo", "metadata": {"model": "gemini-3-pro"}, "reason": "MODEL_CAPACITY_EXHAUSTED"},
+				{"@type": "type.googleapis.com/google.rpc.RetryInfo", "retryDelay": "0.5s"}
+			]
+		}
+	}`)
+	resp := &http.Response{
+		StatusCode: http.StatusServiceUnavailable,
+		Header:     http.Header{},
+		Body:       io.NopCloser(bytes.NewReader(respBody)),
+	}
+
+	params := antigravityRetryLoopParams{
+		ctx:             context.Background(),
+		prefix:          "[test]",
+		account:         account,
+		accessToken:     "token",
+		action:          "generateContent",
+		body:            []byte(`{"input":"test"}`),
+		httpUpstream:    upstream,
+		accountRepo:     repo,
+		isStickySession: true,
+		groupID:         77,
+		sessionHash:     "sticky-503-short",
+		handleError: func(ctx context.Context, prefix string, account *Account, statusCode int, headers http.Header, body []byte, quotaScope AntigravityQuotaScope, groupID int64, sessionHash string, isStickySession bool) *handleModelRateLimitResult {
+			return nil
+		},
+	}
+
+	availableURLs := []string{"https://ag-1.test"}
+
+	svc := &AntigravityGatewayService{cache: cache}
+	result := svc.handleSmartRetry(params, resp, respBody, "https://ag-1.test", 0, availableURLs)
+
+	require.NotNil(t, result)
+	require.NotNil(t, result.switchError)
+	require.True(t, result.switchError.IsStickySession)
+
+	// 验证粘性绑定被清除
+	require.Len(t, cache.deleteCalls, 1)
+	require.Equal(t, int64(77), cache.deleteCalls[0].groupID)
+	require.Equal(t, "sticky-503-short", cache.deleteCalls[0].sessionHash)
+
+	// 验证模型限流已设置
+	require.Len(t, repo.modelRateLimitCalls, 1)
+	require.Equal(t, "gemini-3-pro", repo.modelRateLimitCalls[0].modelKey)
+}
+
+// TestAntigravityRetryLoop_SmartRetryFailed_StickySession_SwitchErrorPropagates
+// 集成测试：antigravityRetryLoop → handleSmartRetry → switchError 传播
+// 验证 IsStickySession 正确传递到上层，且粘性绑定被清除
+func TestAntigravityRetryLoop_SmartRetryFailed_StickySession_SwitchErrorPropagates(t *testing.T) {
+	// 初始 429 响应
+	initialRespBody := []byte(`{
+		"error": {
+			"status": "RESOURCE_EXHAUSTED",
+			"details": [
+				{"@type": "type.googleapis.com/google.rpc.ErrorInfo", "metadata": {"model": "claude-opus-4-6"}, "reason": "RATE_LIMIT_EXCEEDED"},
+				{"@type": "type.googleapis.com/google.rpc.RetryInfo", "retryDelay": "0.1s"}
+			]
+		}
+	}`)
+	initialResp := &http.Response{
+		StatusCode: http.StatusTooManyRequests,
+		Header:     http.Header{},
+		Body:       io.NopCloser(bytes.NewReader(initialRespBody)),
+	}
+
+	// 智能重试也返回 429
+	retryRespBody := `{
+		"error": {
+			"status": "RESOURCE_EXHAUSTED",
+			"details": [
+				{"@type": "type.googleapis.com/google.rpc.ErrorInfo", "metadata": {"model": "claude-opus-4-6"}, "reason": "RATE_LIMIT_EXCEEDED"},
+				{"@type": "type.googleapis.com/google.rpc.RetryInfo", "retryDelay": "0.1s"}
+			]
+		}
+	}`
+	retryResp := &http.Response{
+		StatusCode: http.StatusTooManyRequests,
+		Header:     http.Header{},
+		Body:       io.NopCloser(strings.NewReader(retryRespBody)),
+	}
+
+	upstream := &mockSmartRetryUpstream{
+		responses: []*http.Response{initialResp, retryResp},
+		errors:    []error{nil, nil},
+	}
+
+	repo := &stubAntigravityAccountRepo{}
+	cache := &stubSmartRetryCache{}
+	account := &Account{
+		ID:          17,
+		Name:        "acc-17",
+		Type:        AccountTypeOAuth,
+		Platform:    PlatformAntigravity,
+		Schedulable: true,
+		Status:      StatusActive,
+		Concurrency: 1,
+	}
+
+	svc := &AntigravityGatewayService{cache: cache}
+	result, err := svc.antigravityRetryLoop(antigravityRetryLoopParams{
+		ctx:             context.Background(),
+		prefix:          "[test]",
+		account:         account,
+		accessToken:     "token",
+		action:          "generateContent",
+		body:            []byte(`{"input":"test"}`),
+		httpUpstream:    upstream,
+		accountRepo:     repo,
+		isStickySession: true,
+		groupID:         55,
+		sessionHash:     "sticky-loop-test",
+		handleError: func(ctx context.Context, prefix string, account *Account, statusCode int, headers http.Header, body []byte, quotaScope AntigravityQuotaScope, groupID int64, sessionHash string, isStickySession bool) *handleModelRateLimitResult {
+			return nil
+		},
+	})
+
+	require.Nil(t, result, "should not return result when switchError")
+	require.NotNil(t, err, "should return error")
+
+	var switchErr *AntigravityAccountSwitchError
+	require.ErrorAs(t, err, &switchErr, "error should be AntigravityAccountSwitchError")
+	require.Equal(t, account.ID, switchErr.OriginalAccountID)
+	require.Equal(t, "claude-opus-4-6", switchErr.RateLimitedModel)
+	require.True(t, switchErr.IsStickySession, "IsStickySession must propagate through retryLoop")
+
+	// 验证粘性绑定被清除
+	require.Len(t, cache.deleteCalls, 1, "should clear sticky session in handleSmartRetry")
+	require.Equal(t, int64(55), cache.deleteCalls[0].groupID)
+	require.Equal(t, "sticky-loop-test", cache.deleteCalls[0].sessionHash)
+}