Merge pull request #7139 from HP-network/fix/api-key-model-401-failover

fix: keep API-key accounts available for unknown models

backend/internal/service/model_not_found_error_test.go

@@ -122,6 +122,7 @@ func TestIsOpenAICompatibleModelNotFound400(t *testing.T) {
 		{name: "structured code", body: `{"error":{"code":"model_not_found","message":"No such deployment"}}`, want: true},
 		{name: "unknown provider", body: `{"error":{"message":"Unknown provider for model claude-x"}}`, want: true},
 		{name: "model not found", body: `{"error":{"message":"Model not found: claude-x"}}`, want: true},
+		{name: "unknown model", body: `{"error":{"message":"unknown model: claude-x"}}`, want: true},
 		{name: "model unsupported", body: `{"error":{"message":"The requested model is not supported"}}`, want: true},
 		{name: "plain text compatible gateway", body: `unknown provider for model claude-x`, want: true},
 		{name: "invalid parameter remains terminal", body: `{"error":{"code":"invalid_request_error","message":"Invalid value for temperature"}}`, want: false},

backend/internal/service/openai_gateway_upstream_errors.go

@@ -286,6 +286,10 @@ func (s *OpenAIGatewayService) shouldFailoverOpenAIUpstreamResponse(account *Acc
 }
 
 func isOpenAICompatibleModelNotFound400(respBody []byte) bool {
+	return isOpenAICompatibleModelNotFoundBody(respBody)
+}
+
+func isOpenAICompatibleModelNotFoundBody(respBody []byte) bool {
 	code := strings.TrimSpace(extractUpstreamErrorCode(respBody))
 	if code != "" {
 		return strings.EqualFold(code, "model_not_found")
@@ -296,6 +300,7 @@ func isOpenAICompatibleModelNotFound400(respBody []byte) bool {
 		msg = strings.ToLower(strings.TrimSpace(string(respBody)))
 	}
 	return strings.Contains(msg, "unknown provider for model") ||
+		strings.Contains(msg, "unknown model") ||
 		strings.Contains(msg, "model not found") ||
 		strings.Contains(msg, "model is not supported")
 }

backend/internal/service/ratelimit_service.go

@@ -2468,6 +2468,7 @@ func parseOpenAIImageTryAgainCooldown(body []byte) time.Duration {
 
 const upstreamModelNotFoundCooldown = 30 * time.Minute
 const upstreamModelNotFoundReason = "upstream_404_model_not_found"
+const upstreamModelNotFound401Reason = "upstream_401_model_not_found"
 const upstreamCodexPlanGatedModelCooldown = 30 * time.Minute
 const upstreamCodexPlanGatedModelReason = "upstream_400_codex_plan_gated_model"
 const tempUnschedBodyMaxBytes = 64 << 10
@@ -2493,6 +2494,8 @@ func (s *RateLimitService) HandleUpstreamModelNotFound(ctx context.Context, acco
 	switch {
 	case isUpstreamModelNotFoundError(statusCode, responseBody):
 		cooldown, reason = upstreamModelNotFoundCooldown, upstreamModelNotFoundReason
+	case statusCode == http.StatusUnauthorized && account.Type == AccountTypeAPIKey && account.IsOpenAICompatible() && isOpenAICompatibleModelNotFoundBody(responseBody):
+		cooldown, reason = upstreamModelNotFoundCooldown, upstreamModelNotFound401Reason
 	case isOpenAIOAuthAccount(account) && isOpenAICodexPlanGatedModelError(statusCode, responseBody):
 		cooldown, reason = upstreamCodexPlanGatedModelCooldown, upstreamCodexPlanGatedModelReason
 	default:

backend/internal/service/ratelimit_service_401_test.go

@@ -246,6 +246,29 @@ func TestRateLimitService_HandleUpstreamError_NonOAuth401(t *testing.T) {
 	require.Empty(t, invalidator.accounts)
 }
 
+func TestRateLimitService_HandleUpstreamError_NonOAuthModel401StillDisablesCredentials(t *testing.T) {
+	repo := &rateLimitAccountRepoStub{}
+	service := NewRateLimitService(repo, nil, &config.Config{}, nil, nil)
+	account := &Account{
+		ID:       104,
+		Platform: PlatformOpenAI,
+		Type:     AccountTypeAPIKey,
+	}
+
+	shouldDisable := service.HandleUpstreamError(
+		context.Background(),
+		account,
+		http.StatusUnauthorized,
+		http.Header{},
+		[]byte(`{"error":{"message":"invalid or expired credentials"}}`),
+		"definitely-not-real",
+	)
+
+	require.True(t, shouldDisable)
+	require.Equal(t, 1, repo.setErrorCalls)
+	require.Equal(t, 0, repo.tempCalls)
+}
+
 // TestRateLimitService_HandleUpstreamError_OAuth401DoesNotOverwriteCredentials
 // 回归测试:确保 401 handler 不再使用请求开始时的 account 快照写回 credentials。
 // 原实现会通过 persistAccountCredentials → UpdateCredentials → SetCredentials

backend/internal/service/ratelimit_service_model_not_found_test.go

@@ -69,6 +69,29 @@ func TestRateLimitService_HandleUpstreamError_ModelNotFoundUsesModelRateLimit(t
 	require.WithinDuration(t, time.Now().Add(upstreamModelNotFoundCooldown), call.resetAt, 5*time.Second)
 }
 
+func TestRateLimitService_HandleUpstreamError_APIKeyModel401UsesModelRateLimit(t *testing.T) {
+	repo := &modelNotFoundAccountRepoStub{}
+	svc := &RateLimitService{accountRepo: repo}
+	account := openAIModelNotFoundTempAccount()
+
+	handled := svc.HandleUpstreamError(
+		context.Background(),
+		account,
+		http.StatusUnauthorized,
+		http.Header{},
+		[]byte(`{"error":{"message":"unknown model definitely-not-real"}}`),
+		"definitely-not-real",
+	)
+
+	require.True(t, handled)
+	require.Zero(t, repo.tempCalls)
+	require.Len(t, repo.modelRateLimitCalls, 1)
+	call := repo.modelRateLimitCalls[0]
+	require.Equal(t, account.ID, call.accountID)
+	require.Equal(t, "definitely-not-real", call.scope)
+	require.Equal(t, upstreamModelNotFound401Reason, call.reason)
+}
+
 func TestRateLimitService_HandleUpstreamError_ModelNotFoundWriteFailureDoesNotTempUnschedule(t *testing.T) {
 	repo := &modelNotFoundAccountRepoStub{modelRateLimitErr: errors.New("write failed")}
 	svc := &RateLimitService{accountRepo: repo}